Customized management decision support system for medical institutions

The system integrates and normalizes diverse medical data formats to create a knowledge graph for comprehensive data management and prediction, addressing operational inefficiencies and enhancing decision-making in medical institutions.

KR103046356B1Active Publication Date: 2026-11-10MAX LAB CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
KR1020260099163
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Priority Date
2026-04-09
Filing Date
2026-06-01
Publication Date
2026-11-10
Estimated Expiration
2046-06-01

AI Technical Summary

Technical Problem

Existing medical institution data management systems struggle to integrate and utilize structured and unstructured data from various formats, such as CRM, EMR, and consultation voice/data, leading to incomplete analysis and oversight of contextual information, staff workload, and operational inefficiencies.

Method used

A system that integrates structured and unstructured data through a data collection module, normalization module, medical management ontology construction, data lifecycle flow management, customized prediction, and visualization linkage to create a knowledge graph for managing medical institution operations, including customer and staff management.

Benefits of technology

Enables comprehensive data management and prediction for operational decision-making, reflecting contextual information and staff workload, improving operational efficiency and accuracy in medical institution management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112026066266435-PAT00001_ABST
    Figure 112026066266435-PAT00001_ABST
Patent Text Reader

Abstract

The present invention relates to a management decision support system tailored for medical institutions. More specifically, the invention relates to a system and a method for processing said data that provide information usable for operational decision-making of a medical institution by integrating and collecting structured and unstructured data existing in different formats, such as Customer Relationship Management (CRM) systems, Electronic Medical Record (EMR) systems, appointment management systems, sales management systems, consultation voice data, and handwritten pen chart image data, and by structuring, relating, time-series, predicting, and visualizing said data.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to a management decision support system tailored for medical institutions. More specifically, the invention relates to a system and a method for processing said data that provide information usable for operational decision-making of a medical institution by integrating and collecting structured and unstructured data existing in different formats, such as Customer Relationship Management (CRM) systems, Electronic Medical Record (EMR) systems, appointment management systems, sales management systems, consultation voice data, and handwritten pen chart image data, and by structuring, relating, time-series, predicting, and visualizing said data.

[0002] Furthermore, the present invention may fall under the technical field of not merely storing unstructured records including voice data and pen chart image data generated during the consultation process, but extracting information such as the name of the procedure, the site of administration, the dosage, the purpose of the consultation, expressions related to emotions, and information related to follow-up visits from the records, converting them into structured data, and linking this with a medical management ontology and data lifecycle flow to lead to subsequent prediction and dashboard display.

[0003] Furthermore, the present invention relates to the field of medical information processing technology that can be utilized for consultation operations, procedure switching management, follow-up visit management, personnel management, and management decision support in medical institutions by linking the judgment of next-priority customized proposals at the customer level and schedule control based on burnout status at the medical staff level within the same system. Background Technology

[0004] In general, various data within medical institutions, such as appointment information, treatment information, payment information, consultation notes, handwritten charts, customer inquiry history, and consultation audio, may be managed across different systems or recording methods. For example, appointment-related information may be stored in a CRM system or appointment management system, treatment and procedure-related information recorded in an EMR system, and payment and refund-related information stored in a separate revenue management system. Meanwhile, in the actual consultation setting, information such as potential procedures, customer reactions, highlights by body part, and the need for follow-up consultations is often recorded separately in the form of voice conversations or handwritten pen charts.

[0005] However, as the data dispersed as described above has different formats and recording systems, it can be difficult to perceive the consultation context and procedure transition flow of the same customer as a single continuous dataset. In particular, unstructured records, such as consultation voice data or handwritten pen chart image data, are difficult to utilize directly in subsequent analysis stages even if they contain meaningful information; consequently, some information may be omitted or they may be used only as simple reference materials. Furthermore, since procedure names, body part descriptions, dosage descriptions, or consultation memos written in free-form text do not have consistent notation, there is a possibility that records with the same meaning may be treated as different items during subsequent analysis.

[0006] In such an environment, it can be difficult to identify a consistent relationship structure regarding how a specific customer visited, what skin concerns were repeatedly mentioned during the consultation, whether those concerns actually led to a change in treatment, or how they relate to a return visit after a certain period. Furthermore, if the operational status of a medical institution is judged solely based on the number of appointments or revenue figures, detailed factors such as customer reactions during the consultation, inquiries regarding burdens, follow-up questions, or the daily workload of the medical staff may not be sufficiently reflected.

[0007] Conventionally, technologies could be used to present recommendations based on a customer's past payment history, the time of their last visit, or simply whether a period has expired, or to visualize hospital operational data on a dashboard. However, since these technologies primarily operate based on structured data or simple statistical items, they may have limitations in reflecting contextual information found in unstructured records such as consultation voice recordings and handwritten pen charts, time-series response information, or the processing flow itself, ranging from data generation to prediction and visualization.

[0008] Furthermore, if original source identifiers or consultation session information are severed during the process of speech recognition, optical character recognition, normalization, relationship linking, prediction incorporation, and visualization incorporation following the collection of unstructured data, the data may be treated as unreliable information or omitted during subsequent judgment stages. In such cases, operators may only be able to verify whether the data exists, and it may be difficult to determine to what extent the data has actually been linked and reflected in the analysis.

[0009] Meanwhile, management decision-making at medical institutions can be related not only to recommendation judgments at the customer level but also to the workload of medical staff on the day, movement between spaces, repeated explanations, handling of claims, exceeding schedules, and the status of securing rest. However, conventional reservation management methods may be limited to setting reservation ranges based on working hours or the number of basic slots, making it difficult to accurately reflect the accumulated load conditions at the actual consultation and procedure sites.

[0010] Therefore, a technology configuration may be required that collects, standardizes, and links structured and unstructured data distributed within a medical institution in the same flow, and derives customer-unit relationship information, flow preservation status, next-priority customized proposal judgment information, and medical staff operational status together from the said data to be used for management decision-making. Prior art literature

[65535] Korean Patent Publication No. 10-2019-0135692 The problem to be solved

[0011] The present invention has been devised in consideration of the above points and aims to provide a system capable of linking and collecting reservation data, medical treatment data, sales data, consultation voice data, and handwritten pen chart image data that are distributed inside and outside medical institutions based on the same customer and the same consultation session.

[0012] In addition, the present invention aims to provide a technical configuration capable of extracting information regarding the name of the procedure, the site of the procedure, the dosage administered, the purpose of the consultation, emotional expressions, and information regarding follow-up visits from unstructured records such as the consultation voice data and the handwritten pen chart image data, and converting this into a standardized structured data structure.

[0013] In addition, the present invention aims to provide a medical management ontology construction structure that defines entities and relationships related to customers, inflow channels, skin concerns, procedures, payments, and revisits based on the above-mentioned structured data, and generates relationship strengths that reflect the degree of voice-chart linkage, emphasis, echo utterance, delay time, and modification record history during the consultation process.

[0014] In addition, the present invention aims to provide a structure capable of determining the preservation status of the data lifecycle flow by tracking whether the original source identifier and processing chain are maintained from the stage of original data collection through transformation, normalization, relationship linkage, prediction reflection, and visualization reflection, and by reflecting orphan records, delays between stages, whether automatic recovery is possible, and whether manual intervention is possible.

[0015] In addition, the present invention aims to provide a prediction structure capable of determining acceptance resistance by utilizing the relationship information and the time-series flow data to calculate customer segmentation, revisit patterns, customer churn probability, changes in procedure demand, and next-ranked customized proposal candidates, while reflecting changes in voice response immediately after proposal, periods of low-satisfaction or non-response, inquiries regarding burden, voluntary re-mention, follow-up inquiries, and deviations in the timing of proposal availability per customer.

[0016] In addition, the present invention aims to provide a structure that determines the daily burnout status of medical staff by reflecting equipment load status, number of space switching, repeated explanation time, claim-related voice peak, schedule overrun status, and uninterrupted rest block, and can link the determination result with the display status of available slots, a rest recommendation screen, or an operation review screen.

[0017] Furthermore, the present invention aims to provide a support structure that can be used for management decision-making, such as consultation operations, procedure switching management, follow-up visit management, staffing review, and equipment introduction review, by linking and processing customer-level recommendation judgments and medical staff-level operational judgments within the same system through the components and processing flow described above.

[0018] The problems of the present invention are not limited to those mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0019] In a medical institution-customized management decision support system according to an embodiment of the present invention for solving the above problem,

[0020] A data collection module (100) for collecting structured and unstructured data inside and outside a medical institution;

[0021] An unstructured data normalization module (200) that converts unstructured records collected or converted by the above data collection module (100) into a structured data structure;

[0022] A medical management ontology construction module (300) that creates a data structure in the form of a knowledge graph by defining entities related to medical institution management and relationships between entities based on data standardized by the above-mentioned unstructured data normalization module (200);

[0023] A data lifecycle flow management module (400) that tracks the processing flow of the above structured data, above unstructured data, above structured data, and above knowledge graph data from the collection stage to the transformation, normalization, relationship linkage, prediction reflection, and visualization reflection stages;

[0024] A customized prediction module (500) that generates customer segmentation, revisit pattern analysis, determination of customer churn potential, procedure demand analysis, or next-priority customized proposal data using relationship information generated in the medical management ontology construction module (300) and time series flow data generated in the data lifecycle flow management module (400); and

[0025] A visualization linkage module (600) that provides the results calculated from the above-mentioned customized prediction module (500) in the form of an operation dashboard screen, a report screen, a notification screen, or an operation review screen.

[0026] Includes,

[0027] The data collection module (100) collects unstructured data including consultation voice data and handwritten pen chart image data, and generates text data by performing voice recognition processing on the consultation voice data and optical character recognition processing on the handwritten pen chart image data.

[0028] The above unstructured data normalization module (200) extracts a procedure name, procedure site, dosage, purpose of consultation, price-related expressions, or sentiment-related expressions from the text data and maps them to a standard code system, and

[0029] The above medical management ontology construction module (300) generates relationship information by connecting entities related to customers, inflow channels, skin concerns, procedures, payments, and revisits, and

[0030] The data lifecycle flow management module (400) determines whether the relationship information is maintained along with the original source identifier up to a subsequent stage, and

[0031] The above customized prediction module (500) calculates a predicted value or a recommended value based on the relationship information and the time series flow data, and

[0032] The above visualization linkage module (600) outputs the above predicted value or recommended value together with operational judgment reference information, and

[0033] The above data collection module (100) is linked with at least one of a customer relationship management system, an electronic medical record system, a reservation management system, a sales management system, a wearable badge unit, and a handwritten pen chart input means to collect reservation data, sales data, consultation voice data, and pen chart image data, and

[0034] From the above reservation data, the reservation date and time, reservation channel, reservation procedure name, reservation manager, and reservation change history are obtained, and

[0035] From the above sales data, payment date and time, payment amount, sales items by procedure, refund status, and discount information are obtained,

[0036] The above consultation voice data is linked with the consultation start time, consultation end time, consultation manager identification information, customer identification information, and consultation session identification information, and

[0037] The above pen chart image data is linked with the time of shooting, author identification information, customer identification information, and consultation session identification information, and

[0038] At least one of source identification information, collection time information, customer identification information, consultation identification information, and medical treatment identification information is assigned to each of the above-mentioned collected data to enable linkage of identical data between subsequent modules, and

[0039] Speech recognition processing is performed on the above consultation voice data to generate a voice transcription, and optical character recognition processing is performed on the above pen chart image data to generate a character extraction result.

[0040] The above original file and the above voice transcription or the above character extraction result are stored in a form that allows for mutual mapping, and

[0041] It checks for missing required items, file corruption, or duplicate collection, saves the check results as status values ​​indicating normal, suspected missing, candidate duplicate, awaiting conversion, or need for re-collection, and provides them as criteria for subsequent processing,

[0042] The above unstructured data normalization module (200) has, for the voice transcription, optical character recognition result, free input memo, and counselor input text received from the data collection module (100),

[0043] A preprocessing step that performs at least one of spacing correction, special character removal, abbreviation expansion, synonym integration, typographical correction, and unit expression unification;

[0044] An item extraction step for extracting the procedure name, procedure site, dosage, purpose of consultation, expression of dissatisfaction, expression of inducing a revisit, price-related expression, and sentiment-related expression from the above-mentioned preprocessed text;

[0045] A code mapping step for mapping the above extracted items to standard procedure codes, site codes, dosage codes, sentiment classification codes, or price inquiry codes.

[0046] Perform,

[0047] In cases where multiple procedure names and multiple sites are listed together in a single sentence or multiple text fragments within the same consultation session, parsing rules or a learning model are used to separate the correspondence between the procedure, the site, and the dosage.

[0048] Voice transcripts, pen chart records, and consultation notes belonging to the same consultation session are merged and reconstructed into a single integrated structured record, and

[0049] By referring to an internal dictionary containing procedure abbreviations, equipment names, part expressions, and dosage expressions used within the medical institution, and normalizing multiple expressions with the same meaning into the same standard item,

[0050] Information distributed across different media is used as consistent input values ​​in the subsequent ontology construction and prediction stages, and

[0051] The above medical management ontology construction module (300) defines customers, counselors, medical staff, visit route, skin concerns, procedure name, procedure site, dosage, price, consultation sentiment, reservation conversion status, and revisit cycle as entities or attributes, and

[0052] Based on the above customer identification information, consultation identification information, procedure identification information, and occurrence time information, records generated in the reservation stage, consultation stage, treatment stage, payment stage, and follow-up visit stage are connected in chronological order to form a knowledge graph path, and

[0053] After first assigning default weights to the relationship between a specific skin concern entity and a specific treatment entity,

[0054] The number of overlapping occurrences representing the number of matching events in which the same procedure is identified together in voice transcriptions and pen chart records within the same consultation session,

[0055] Highlighting marking density calculated by dividing the number of underline, asterisk, circle, or box mark objects formed in the area where a specific procedure is recorded in the above pen chart by the normalized area of ​​said area,

[0056] The number of echo speeches, representing the number of times the same procedure name or expression of the same efficacy appears repeatedly in customer speech within a preset time interval immediately after the medical staff's explanation.

[0057] A synchronization delay time representing the difference between the time when a specific procedure was first mentioned in the medical staff's verbal explanation and the time when said procedure was first indicated on the chart, and

[0058] Number of revision entries, including overwriting, duplicate marking, and re-entry after deletion, for the same chart area

[0059] By collecting,

[0060] The basic weight is adjusted upward based on the number of overlapping occurrences, the emphasis marking density, and the number of echo utterances, and the basic weight is adjusted downward based on the synchronization delay time and the number of modification entries.

[0061] The above correction result is stored as a dynamic relationship weight, which is an attribute value of a knowledge graph edge connecting the skin concern entity and the treatment entity, and the above dynamic relationship weight is provided as an input value to distinguish between priority review relationships and subordinate relationships in a subsequent prediction step.

[0062] The above medical management ontology construction module (300) corrects the dynamic relationship weights,

[0063] If the above synchronization delay time is within the preset first threshold value, it is determined that the voice description and chart display are linked within the same consultation flow, and the above basic weight is increased by the first increment range, and

[0064] If the number of overlapping occurrences is greater than or equal to a preset second threshold value and simultaneously the emphasis marking density is greater than or equal to a preset third threshold value, a greater synergy correction is performed than when a single condition is satisfied, and

[0065] As the number of echo utterances increases, the weight is added stepwise, and

[0066] If the above synchronization delay time exceeds the above first reference value, the reduction range is gradually expanded according to the degree of excess, and

[0067] If the number of modification entries mentioned above is greater than or equal to the preset fourth threshold value, it is determined that there is repeated modification regarding the same chart area, and the weight is further lowered.

[0068] If the chart display time precedes the voice mention time and the synchronization delay time is calculated as a negative value, it is determined as a pre-written template event and excluded from the synergy correction target.

[0069] Range limit control is performed so that the above weights do not exceed a preset upper limit, thereby reducing situations where the relationship strength is excessively biased by specific input values, and

[0070] The resulting edge attribute values ​​are provided as criteria to distinguish between priority and subordinate relationships in the subsequent prediction step, and

[0071] The above data lifecycle flow management module (400) assigns an original source identifier to each data, including at least one of an original voice file identifier, an original image file identifier, a customer identifier, a consultation session identifier, and a conversion task identifier.

[0072] The source chain completeness ratio is calculated using the ratio of the number of processing steps actually connected without disconnection among the collection step, text transformation step, normalization mapping step, ontology linkage step, prediction reflection step, and visualization reflection step to the total number of processing steps, and

[0073] The number of nodes among the data nodes that have reached the final prediction reflection stage or the visualization reflection stage in which the above original source identifier is maintained without interruption is calculated as the number of valid nodes, and

[0074] Records that are not mapped to customer identification information, consultation session identification information, or original source identifier are identified as orphan records, and

[0075] Calculate a discontinuous timestamp accumulation value by accumulating the excess portion of the interval exceeding a preset allowable threshold among the time differences between adjacent processing steps, and

[0076] In the event of partial omission of identifiers or failure of inter-stage connections, the number of times recovery is performed according to automatic recovery rules is calculated as the number of automatic recovery successes, and the number of times an operator directly reassigns identifiers or modifies mapping relationships is calculated as the number of manual intervention corrections,

[0077] If the number of valid nodes is greater than or equal to a preset valid node threshold and simultaneously the source chain integrity ratio is greater than or equal to a preset chain integrity threshold, the data flow preservation status is classified upwardly by a larger margin than when a single condition is satisfied, and

[0078] If the above orphan record occurs or the above timestamp discontinuous accumulation value exceeds a preset discontinuous accumulation threshold value, the above flow preservation status is classified downward, provided that if the number of successful automatic recoverys is greater than the number of manual intervention corrections, the downward range is limited.

[0079] By ensuring that the data set with a high flow preservation state is prioritized for input into the customized prediction module (500) and that the data set with a low flow preservation state is moved to a reprocessing queue or output with a warning indication through the visualization linkage module (600),

[0080] Data with its original source preserved and data with broken or missing sources can be treated differently in subsequent processing steps. Effects of the invention

[0081] According to the present invention, structured and unstructured data distributed within and outside a medical institution can be linked and collected based on the same customer and the same consultation session, so that the flow from reservation, consultation, treatment, payment, and follow-up visit can be utilized to manage the entire process as a single data system.

[0082] In addition, according to the present invention, information regarding the name of the procedure, the site of the procedure, the dosage administered, the purpose of the consultation, emotional expressions, and information regarding follow-up visits can be extracted from unstructured records, such as consultation voice data and handwritten pen chart image data, and converted into standardized structured data. Therefore, consultation information that was recorded in different formats can be utilized to be used consistently in subsequent analysis steps.

[0083] In addition, according to the present invention, relationships are formed with entities related to customers, inflow channels, skin concerns, procedures, payments, and revisits based on structured data, and since the degree of voice-chart linkage, degree of emphasis, echo utterance, delay time, and modification record history during the consultation process can be reflected together with respect to said relationships, they can be used to generate relationship information that is distinct from simple keyword co-occurrence.

[0084] In addition, according to the present invention, since the preservation status of the data lifecycle flow can be determined by reflecting the original source identifier, processing step coverage, occurrence of orphan records, timestamp discontinuity, and automatic recovery status together, it can be utilized to manage the status of data chain maintenance used in subsequent prediction or visualization steps.

[0085] In addition, according to the present invention, customer segmentation, revisit patterns, customer churn probability, changes in procedure demand, and next-priority customized proposal candidates can be calculated using relationship information and time-series flow data, and acceptance resistance status can be determined by reflecting the voice response immediately after the proposal and the deviation in proposal timing by customer, so it can be utilized to adjust proposal priority or hold.

[0086] In addition, according to the present invention, the daily burnout status of medical staff can be determined by reflecting together the equipment load status, the number of space switchings, the repeated explanation time, the claim-related voice peak, the schedule overrun status, and the uninterrupted rest block, so it can be utilized to adjust the display status of available slots for reservation or to inform the operator of the need for rest.

[0087] In addition, according to the present invention, since recommendation judgments at the customer level and schedule control judgments at the medical staff level can be processed in conjunction within the same system, it can be used to support medical institution management decision-making, including consultation operation, procedure switching management, follow-up visit management, personnel placement review, and provision of operation screens.

[0088] In addition, according to the present invention, by reflecting the processing flow from the collection of unstructured data to prediction and visualization together with customer responses and the status of medical staff, a basis can be provided for a medical institution operator to review multiple judgment criteria within a single screen or a single system flow.

[0089] The effects according to the present invention are not limited to those exemplified above, and a wider variety of effects are included within the present invention. Brief explanation of the drawing

[0090] Figure 1 illustrates a flowchart between all components according to the present invention. Figure 2 illustrates a flowchart of the dynamic relationship weight processing logic within a knowledge graph according to the present invention. Figure 3 illustrates a flowchart of the evaluation logic for preventing the evaporation of unstructured data lifecycle according to the present invention. Figure 4 illustrates a flowchart of the logic for determining resistance to accepting a next-ranked customized proposal based on dynamic behavior according to the present invention. Figure 5 illustrates a flowchart of the medical staff burnout prediction and schedule control logic according to the present invention. Specific details for implementing the invention

[0091] Hereinafter, various embodiments are described in more detail with reference to the attached drawings. The embodiments described in this specification may be modified in various ways. Specific embodiments may be depicted in the drawings and described in detail in the detailed description. However, specific embodiments disclosed in the attached drawings are intended only to facilitate understanding of various embodiments. Accordingly, the technical concept is not limited by specific embodiments disclosed in the attached drawings, and it should be understood that it includes all equivalents or substitutions that fall within the spirit and scope of the invention.

[0092] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but these components are not limited by the aforementioned terms. The aforementioned terms are used solely for the purpose of distinguishing one component from another.

[0093] In this specification, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof. When a component is described as being “connected” or “connected” to another component, it should be understood that it may be directly connected to or connected to that other component, or that there may be other components in between. On the other hand, when a component is described as being “directly connected” or “directly connected” to another component, it should be understood that there are no other components in between.

[0094] Meanwhile, a "module" or "part" for a component as used in this specification performs at least one function or operation. Furthermore, a "module" or "part" may perform a function or operation by hardware, software, or a combination of hardware and software. Additionally, a plurality of "modules" or a plurality of "parts," excluding a "module" or "part" that must be performed on specific hardware or on at least one processor, may be integrated into at least one module. A singular expression includes a plural expression unless the context clearly indicates otherwise.

[0095] The communication method, storage method, processor driving method, database management method, and screen output method for performing data collection, text conversion, normalization, ontology linkage, prediction reflection, and visualization reflection described in this specification are described within the scope combined with the judgment criteria and processing flow of the present invention. However, known information processing means such as general network communication, file storage, database query, user authentication, screen rendering, and log recording for driving each of the above modules may be applied within the scope that a person skilled in the art can ordinarily implement, and the following description focuses on the data linkage structure, relationship weight correction structure, flow preservation state determination structure, and operation judgment linkage structure of the present invention. Furthermore, regarding operation embodiments and configurations described broadly without detailed explanation below, they follow the prior art and are omitted in order to focus on describing the purpose of the present invention and the resulting effects.

[0096] Furthermore, in describing the present invention, if it is determined that a detailed description of related known functions or configurations may unnecessarily obscure the essence of the invention, such detailed description is abbreviated or omitted.

[0097] The medical institution customized management decision support system according to the present invention may include a data collection module (100), an unstructured data normalization module (200), a medical management ontology construction module (300), a data lifecycle flow management module (400), a customized prediction module (500), and a visualization linkage module (600).

[0098] In this specification, the logic for processing dynamic relationship weights within the knowledge graph of the medical management ontology construction module (300) illustrated in FIG. 2 is hereinafter referred to as “Process 1.” Additionally, the logic for determining the preservation status of unstructured data lifecycle flow of the data lifecycle flow management module (400) illustrated in FIG. 3 is hereinafter referred to as “Process 2.” Additionally, the logic for determining the resistance status to accepting the next-ranked customized proposal of the customized prediction module (500) illustrated in FIG. 4 is hereinafter referred to as “Process 3.” Furthermore, the logic for predicting the burnout status of medical staff and controlling the schedule of the customized prediction module (500) and the visualization linkage module (600) illustrated in FIG. 5 is hereinafter referred to as “Process 4.” Process 1 and Process 2 above may correspond to a structure for determining data relationality and data flow preservation status, which is primarily defined in the claims of the present application, and Process 3 and Process 4 above may be described as an embodiment for applying the relational information and flow preservation status generated in Process 1 and Process 2 to a subsequent prediction and operation screen. A data collection module (100) according to an embodiment of the present invention may collect structured data and unstructured data used for management decision-making from multiple data sources inside and outside a medical institution. The data collection module (100) may acquire reservation data, sales data, consultation voice data, and pen chart image data by linking with a customer relationship management system, an electronic medical record system, a reservation management system, a sales management system, a wearable badge unit, and a handwritten pen chart input means. The data collection module (100) may not simply receive individual files or individual records separately, but may assign common identification information from the collection stage so that they can be mutually linked in subsequent stages based on the same customer or the same consultation session.

[0099] The data collection module (100) can collect data through at least one of application program interface linkage, database query, file upload, message queue-based reception, or log collection methods according to the interface structure of an external system. For example, it can receive the consultation reservation date and time, reservation channel, reservation procedure name, consultation manager, reservation change history, etc. from the customer relationship management system, and can receive the treatment date and time, procedure history, procedure item, payment item, and customer basic history information from the electronic medical record system. In addition, it can obtain the payment amount, discount application status, refund status, sales item by procedure, and payment time information from the sales management system.

[0100] The data collection module (100) can also collect unstructured data. Specifically, it can receive a voice file linked to the consultation start time and consultation end time from the wearable badge unit, and the voice file may be linked with consultation manager identification information and customer identification information. In addition, the data collection module (100) can acquire handwritten pen chart image data that is captured or uploaded after the consultation ends, and the pen chart image data may be assigned the time of capture, author identification information, customer identification information, and consultation session identification information. The data collection module (100) can store the original file itself and metadata together so that the voice data and image data can be directly used by a subsequent processing module.

[0101] The data collection module (100) may further include a preprocessing function that converts collected unstructured data into text data. For example, the data collection module (100) may generate a voice transcription by performing voice recognition processing on consultation voice data, and may generate a character extraction result by performing optical character recognition processing on the handwritten pen chart image data. In this case, the data collection module (100) may store the original voice file and the voice transcription, and the original image and the optical character recognition result, in a form that allows mutual mapping. Accordingly, in subsequent steps, normalization and analysis can be performed while maintaining the correspondence between the original data and the conversion result.

[0102] The data collection module (100) can check the data status during the collection process. For example, it can determine whether essential items are missing, whether the file is corrupted, whether collection has failed, whether duplicate data has been received, or whether there is a discrepancy in identification information. The results of the check can be stored as status values ​​such as normal, suspected missing, candidate for duplicate, waiting for conversion, or need for re-collection. These status values ​​can be used as criteria for a subsequent module to determine whether to process the data immediately, make it subject to re-inspection, or send it to an exception handling flow.

[0103] As an example of the above data collection module (100), a specific customer may book an initial consultation through a social network service advertisement. The above data collection module (100) may collect the customer’s reservation history and inflow channel information from a customer relationship management system, collect the customer’s medical history from an electronic medical record system, acquire a consultation voice file from a wearable badge unit, and receive a handwritten pen chart image uploaded after the consultation ends. Subsequently, the above data collection module (100) may perform voice recognition processing on the voice file and perform optical character recognition processing on the pen chart image to form a basic data set linked based on the same customer and the same consultation session.

[0104] An unstructured data normalization module (200) according to one embodiment of the present invention can convert unstructured records, such as voice transcriptions, optical character recognition results, free-form memos, counselor input text, and handwritten chart records collected or converted by a data collection module (100), into a structured data structure. The unstructured data normalization module (200) does not store the text written in a free-form sentence as is, but can decompose the procedure name, procedure site, dosage, purpose of consultation, expression of dissatisfaction, expression of inducing a revisit, price-related expression, and emotion-related expression included in the text and reconstruct them into standardized items.

[0105] The above-described unstructured data normalization module (200) can analyze the unstructured records using natural language processing, a large language model, a term dictionary, a rule-based parser, or a combination thereof. The analysis may first begin with a text preprocessing step, in which at least one of spacing correction, removal of special characters, abbreviation expansion, synonym integration, error correction, unit expression standardization, and removal of unnecessary notations may be performed. For example, variously written procedure expressions such as “Botok,” “Botox,” “Jaw Botok,” and “Jaw Botox” can be normalized into the same standard procedure name candidate.

[0106] The above-described unstructured data normalization module (200) can extract individual items from preprocessed text. For example, the above-described unstructured data normalization module (200) can identify expressions related to the name of the procedure, the treatment area, the dosage, the desired timing, the purpose of consultation, the price inquiry, concerns about side effects, or the intention to revisit within the text. Additionally, when multiple names of procedures and multiple areas are listed together in a single sentence, the above-described unstructured data normalization module (200) can separate the correspondence between the procedure, the area, and the dosage using parsing rules or a learning model. For example, when the expression “Thermage 600 shots forehead between eyebrows” is input, the above-described unstructured data normalization module (200) can separate the name of the procedure into “Thermage,” the dosage into “shots,” and the treatment areas into “forehead” and “between eyebrows.”

[0107] The above-described unstructured data normalization module (200) can map extracted items to a standard code system. For example, procedure names can be converted into standard procedure codes, area information into area codes, dosage expressions into dosage codes, consultation sentiment expressions into sentiment classification codes, and price-related expressions into price inquiry codes. The above-described standard code system can be corrected by referring to an internal dictionary containing abbreviations, equipment names, area expressions, and dosage expressions frequently used within the medical institution. Accordingly, multiple expressions having the same meaning can be treated as the same item in a subsequent analysis step.

[0108] The above-described unstructured data normalization module (200) can reconstruct a single structured record by merging multiple text fragments, not just a single text fragment. For example, if a voice transcript of a consultation includes the name of the procedure and the purpose, a handwritten pen chart indicates the area and dosage, and a counselor's memo additionally records a price inquiry, the above-described unstructured data normalization module (200) can generate a single integrated structured record by merging the three fragments based on the same consultation session. Accordingly, information that was scattered across different media can be used as a consistent input value in the subsequent ontology construction and prediction stages.

[0109] As an example of the above-mentioned unstructured data normalization module (200), there may be a case where the voice transcript of the consultation includes the expression “I would like a consultation on the forehead area as well after the jaw Botox last time,” and the handwritten pen chart contains abbreviated entries such as “forehead,” “between eyebrows,” and “shot.” ​​The above-mentioned unstructured data normalization module (200) can extract “jaw” and “forehead” as area items from the voice transcript and identify “Botox” as a procedure item, and by merging with the pen chart record, it can finally generate a structured record including the consultation target area, past procedure history, new consultation desired area, and procedure conditions.

[0110] A medical management ontology construction module (300) according to one embodiment of the present invention defines entities related to medical institution management and relationships between entities based on data standardized by an unstructured data normalization module (200), and can organize this into a data structure in the form of a knowledge graph. The medical management ontology construction module (300) can define entities or attributes such as customers, counselors, medical staff, visit paths, skin concerns, procedure names, procedure sites, dosages, prices, consultation sentiment, reservation conversion status, and revisit cycles, and can connect relationships between the entities according to a continuous event flow.

[0111] The above medical management ontology construction module (300) can first define entity types and relationship types. For example, a “inflow” relationship can be defined between the customer entity and the visit path entity, a “pleading” relationship can be defined between the customer entity and the skin concern entity, and a “recorded as consultation result” relationship can be defined between the counselor entity and the consultation sentiment entity. Additionally, a “applied” relationship can be established between the procedure entity and the body part entity, a “paid” relationship between the customer entity and the payment entity, and a “revisited” relationship between the customer entity and the revisit entity.

[0112] The above medical management ontology construction module (300) can connect records generated at different stages based on the same customer or the same consultation session. For example, records generated at the reservation stage, consultation stage, treatment stage, payment stage, and revisit stage can be connected in chronological order to represent the procedure transition flow or revisit flow as a graph path. In this case, since a creation time or occurrence time can be assigned to each relationship, subsequent modules can refer not only to whether a simple relationship exists but also to the chronological order of the relationships.

[0113] The medical management ontology construction module (300) can create entities using standard procedure codes, site codes, dosage codes, sentiment codes, price information, and consultation result information received from the unstructured data normalization module (200), and can connect reservation channels, visit timestamps, consultation staff, medical staff, payment items, and customer basic history information received from the data collection module (100) as additional attributes. Relationships between the entities can be created based on identical customer identification information, identical consultation identification information, or identical procedure identification information.

[0114] The medical management ontology construction module (300) can generate relationship feature values ​​by identifying patterns that appear repeatedly on the relationship network. For example, if a group of customers introduced from a specific inflow channel repeatedly complains of specific skin concerns and tends to revisit within a certain period after switching to the same procedure, the medical management ontology construction module (300) can store this as a relationship pattern that can be referenced in a subsequent prediction step. The relationship feature values ​​can be used as input characteristics for segmentation, judgment of conversion possibility, or judgment of revisit possibility in a subsequent customized prediction module (500).

[0115] As an example of the above medical management ontology construction module (300), there is a case where a customer who entered through a social network service advertisement receives a consultation for frown line improvement, switches to a Botox procedure, and then receives a follow-up consultation after 4 months. The above medical management ontology construction module (300) sequentially connects the customer entity, the inflow channel entity, the skin concern entity, the counselor entity, the procedure entity, the payment entity, and the follow-up entity, and can store the time of occurrence and sales information for each stage as additional attributes. The relationship information configured in this way can be used to analyze how a combination of a specific inflow channel and a specific skin concern is connected to actual procedure conversion and follow-up visits.

[0116] A data lifecycle flow management module (400) according to one embodiment of the present invention can track and manage the processing flow from the point in time when consultation voice data, handwritten pen chart image data, free text records, reservation data, and sales data are generated to normalization, storage, linkage, analysis, and visualization. To reduce situations where unstructured data is lost on a one-time basis or omitted in subsequent analysis steps, the data lifecycle flow management module (400) can record at least one of the point in time of generation, point in time of transformation, point in time of storage, point in time of completion of normalization, point in time of relationship linkage, point in time of prediction reflection, and point in time of visualization reflection for each data.

[0117] The data lifecycle flow management module (400) can manage original voice files, voice transcripts, optical character recognition result texts, normalized records, ontology relationship information, prediction input records, and dashboard output records as a single flow chain by connecting them step by step. The flow chain can be formed based on identical customer identification information, identical consultation session identification information, original file identifiers, and processing stage identification information. Accordingly, it is possible to track which stage specific data was generated, which stage it was transformed, and which stage it was omitted or merged.

[0118] The data lifecycle flow management module (400) can generate time-series flow data by arranging multiple events belonging to the same customer or the same consultation session in chronological order. For example, the consultation start time, voice file creation time, consultation end time, pen chart image capture time, voice recognition completion time, optical character recognition completion time, normalization completion time, ontology linkage time, and prediction reflection time can be arranged in chronological order. The time-series flow data can be used by a subsequent module to interpret the progress status of a specific consultation case as a continuous flow.

[0119] The data lifecycle flow management module (400) can assign a status value to manage missing events or duplicate records. For example, if the original file exists but the transformation result is not generated, a transformation missing status can be assigned, and if duplicate records occur at the same stage, a merge candidate status can be assigned. Additionally, if the normalization result exists but is not linked to customer identification information, a connection missing status can be assigned. The status value can be used as a standard for subsequent reprocessing, manual inspection, or automatic merging procedures.

[0120] The data lifecycle flow management module (400) can receive original file metadata, collection logs, and transformation logs from the data collection module (100), receive text before and after normalization, standard code mapping results, and exception processing results from the unstructured data normalization module (200), and receive entity connection results and relationship creation time information from the medical management ontology construction module (300). Accordingly, the data lifecycle flow management module (400) can identify which processing stage each data is located in step by step.

[0121] As an example of the data lifecycle flow management module (400) described above, a voice file is generated from a wearable badge unit at the start of a consultation, a handwritten pen chart image is captured after the consultation ends, and subsequently, a voice recognition result and an optical character recognition result are generated sequentially. The data lifecycle flow management module (400) can connect each of the above steps into a single consultation flow by assigning a timestamp and a status value. Accordingly, it is possible to distinguish and manage whether a specific consultation case has actually reached the prediction input record, has been held in the normalization stage, or has been reflected in the visualization stage.

[0122] A customized prediction module (500) according to one embodiment of the present invention can calculate a predicted value or a recommended value related to the operation of a medical institution using relationship information generated in a medical management ontology construction module (300) and time-series flow data generated in a data lifecycle flow management module (400). The customized prediction module (500) can perform at least one of customer segmentation, revisit cycle pattern analysis, determination of potential customer churn, analysis of changes in procedure demand, and generation of next-priority customized proposal data.

[0123] The customized prediction module (500) can first classify customer groups based on inflow paths, consultation content, treatment history, revenue contribution, revisit history, sentiment classification results, reservation cancellation history, refund history, or complaint keyword appearance history. Subsequently, the customized prediction module (500) can analyze the revisit interval, treatment conversion frequency, churn frequency, and increase / decrease trends in demand for specific treatments within the same customer group or customer groups with similar relationship patterns. Based on the analysis results, the customized prediction module (500) can calculate the need for follow-up consultation, the possibility of revisiting, the need for churn management, or candidates for treatments of interest for a specific customer or a specific customer group.

[0124] The customized prediction module (500) can receive relationship information between entities, relationship frequency information, relationship path information, and customer-specific connection pattern information from the medical management ontology construction module (300). Additionally, it can receive time-series flow data including the consultation occurrence time, procedure conversion time, payment time, and revisit time from the data lifecycle flow management module (400). The customized prediction module (500) can combine the relationship information and the time-series flow data to analyze which path leads a specific customer group to procedure conversion or churn.

[0125] The above-described customized prediction module (500) can generate next-rank customized proposal data using the above analysis results. The above-described next-rank customized proposal data may include a recommended procedure name, a recommendation time, a recommended target customer group, a recommendation basis item, and a priority review ranking. For example, among a customer group that visited through a specific inflow channel, if a customer group that mentions a specific skin concern in the early part of the consultation session and has a history of re-consultation within 3 months after the first procedure shows a relatively high tendency to revisit, the above-described customized prediction module (500) can generate next-rank customized proposal data including a specific procedure or package proposal at the time of subsequent consultation.

[0126] In addition, the customized prediction module (500) can generate a churn management target indicator for customer groups in which an increase in the number of reservation cancellations, a decline in consultation sentiment, a history of refunds, and an increase in the interval between revisits are observed together. Accordingly, the customized prediction module (500) can perform not only a simple recommendation function but also a function to separately identify customer groups that require operational attention.

[0127] As an example of the above-mentioned customized prediction module (500), a case may be cited in which a group of customers who visited through a specific inflow channel during the last 6 months mention concerns regarding frown lines or lifting in the early part of the consultation session and have a record of re-consultation within 3 months after the procedure are classified into a separate segment.

[0128] A visualization linkage module (600) according to one embodiment of the present invention may provide the results calculated by the customized prediction module (500) in the form of a dashboard screen, a report screen, a notification screen, or an operation review screen so that a medical institution operator can check them. The visualization linkage module (600) may comprehensively display next-ranked customized proposal data, a human resource scorecard, marketing indicators, indicators of changes in procedure demand, indicators of revisit patterns, and information on churn management targets.

[0129] The visualization linkage module (600) can classify the result values ​​transmitted from the customized prediction module (500) by item and output them in the form of a time series chart, comparison table, customer group summary table, performance table by counselor, trend graph by procedure, or warning display. Additionally, the visualization linkage module (600) can provide filtered results based on a period, inflow channel, counselor, procedure group, customer group, or operation item according to the user's selection.

[0130] The visualization linkage module (600) can receive the creation time and reflection status of a specific record together from the data lifecycle flow management module (400). Accordingly, the visualization linkage module (600) can display the data's recency, whether reflection is complete, or whether inspection is required on the screen. For example, by displaying the data reflection time or inspection caution status together next to a specific demand forecast result, the operator can determine when to use the result.

[0131] The visualization linkage module (600) can provide a reference screen for operational decisions by combining sales data, reservation data, consultation data, and prediction results. For example, the visualization linkage module (600) can display consultation conversion rates, procedure conversion rates, revisit rates, average sales contribution per counselor, and demand change trends by procedure group together on a single dashboard for a specific monthly period. Additionally, it can provide a screen for reviewing whether to introduce new equipment by displaying the increase in reservations and sales contribution of procedures related to specific equipment together, and can also provide a screen for reviewing whether to adjust the operational schedule by comparing the reservation concentration period before and after holidays with the availability period of counseling personnel.

[0132] As an example of the above-mentioned visualization linkage module (600), the consultation conversion rate and actual payment conversion rate of a customer group that entered through a specific inflow channel are displayed together on the monthly operation dashboard screen, and the trend of demand change for the relevant treatment group over the past three months is displayed overlaid. Based on the above results, the above-mentioned visualization linkage module (600) can provide reference information so that the operator can review whether to allocate the marketing budget, deploy consultation personnel, or expand a specific treatment group.

[0133] A data collection module (100) according to one embodiment of the present invention can collect structured data and unstructured data inside and outside a medical institution and transmit them to an unstructured data normalization module (200). The unstructured data normalization module (200) can convert text-based unstructured records into standardized structured data and transmit them to a medical management ontology construction module (300) and a data lifecycle flow management module (400). The medical management ontology construction module (300) can generate a knowledge graph by defining relationships between the structured data, and the data lifecycle flow management module (400) can manage the flow from data generation to analysis reflection in a time-series manner. The customized prediction module (500) can generate predicted values ​​or recommended values ​​using the knowledge graph and the time-series flow data, and the visualization linkage module (600) can provide the results to an operation dashboard screen or a review screen.

[0134] A medical management ontology construction module (300) according to one embodiment of the present invention can calculate dynamic relationship weights representing the strength of the relationship between a specific skin concern entity and a specific procedure entity by mutually linking voice data and pen chart image data generated during the consultation process. The medical management ontology construction module (300) can adjust the relationship weights by not merely connecting relationships based on whether a specific procedure name exists in the text, but by also considering in what context the procedure is mentioned within the same consultation session, in what form it is displayed in the pen chart, and in what response the customer re-mentioned the procedure or its efficacy.

[0135] The above medical management ontology construction module (300) may collect multiple judgment criteria for calculating the relationship weights. The judgment criteria may include: a number of overlapping occurrences representing the number of matching events in which the same procedure is identified together in voice transcriptions and pen chart records within the same consultation session; a highlighting marking density calculated by dividing the number of highlighting objects, such as underlines, stars, circles, and boxes, formed in the area where a specific procedure is written in the pen chart by the normalized area of ​​the said area; a synchronization delay time representing the difference between the time when a specific procedure is first mentioned in the medical staff's verbal explanation and the time when the procedure is first displayed on the pen chart; an echoing utterance count representing the number of times the same procedure name or expression of the same efficacy appears repeatedly in customer speech within a preset time interval immediately after the medical staff's explanation; and a number of modification entries including overwriting, duplicate marking, and re-writing after deletion for the same chart area.

[0136] The number of overlapping occurrences mentioned above can be calculated based on the number of events in which voice transcriptions and pen chart records correspond to each other in the same consultation session. For example, if a specific procedure name is mentioned in the voice transcription and the same procedure entry or highlight object in the pen chart is identified within a preset time window before and after the mention, the medical management ontology construction module (300) can determine this as one overlapping occurrence. The number of echo utterances mentioned above can be tallied only for the customer utterance section after speaker separation processing, and if the customer mentions the same procedure name or the same efficacy expression again within a preset time window after the medical staff utters a specific procedure name or efficacy expression, it can be counted as one occurrence.

[0137] The above-mentioned emphasis marking density can be calculated based on a text area in which a specific procedure name is written or a display area associated with the procedure name in the above-mentioned pen chart image. The above-mentioned medical management ontology construction module (300) can extract a bounding box or a normalized pixel area of ​​the above-mentioned area and calculate the emphasis marking density using at least one of the number of underline, asterisk, circle mark, and box mark objects identified within the above-mentioned area and the pixel area occupied by said objects.

[0138] In one embodiment, the emphasis marking density may be calculated by pre-setting a reference number of emphasis objects per 100 units of normalized area and converting the actual number of detected emphasis objects into a normalized value in the range of 0 to 1 by comparing it with the reference number of emphasis objects. For example, if the reference number of emphasis objects per 100 units of normalized area is set to 5, the emphasis marking density may be calculated as 0.6 when 3 emphasis objects are detected within the same area, and as 0.8 when 4 emphasis objects are detected. In this case, if the reference value for the emphasis marking density is set to 0.5, a case where approximately 3 or more emphasis objects are detected per 100 units of normalized area can be determined as a state where the criteria are satisfied.

[0139] In another embodiment, the emphasis marking density may be calculated based on an area ratio obtained by dividing the total pixel area of ​​the emphasized object within the area by the total pixel area of ​​the area. In this case, the medical management ontology construction module (300) stores the object count-based density and the area ratio-based density as different sub-indicators, and can determine that the emphasis marking is formed above the standard if either sub-indicator is greater than or equal to a preset standard value or if the combined value of the two sub-indicators is greater than or equal to a preset standard value. The emphasis marking density may be used as a normalized value in the range of 0 to 1 throughout this specification. The synchronization delay time may be calculated using time information of voice data and chart data. The medical management ontology construction module (300) can identify the timestamp at the time when a specific procedure was first mentioned in the voice data acquired from the wearable badge unit. Additionally, the medical management ontology construction module (300) can identify the time when the procedure was first displayed by using at least one of a digital pen input log, an electronic chart editing log, a pen chart image frame by time, or a chart saving time log. For example, if a specific procedure is first mentioned in the voice data at 14:05:10 and the same procedure is identified as first displayed in the electronic chart editing log at 14:05:15, the synchronization delay time can be calculated as 5 seconds. In a still image-based environment where no separate editing log exists, the saving time or upload time of the pen chart image can be used as auxiliary time information.

[0140] The number of modification entries mentioned above can be calculated based on the number of multiple modification events for the same chart area. The modification events may include overwriting within the same text area, rewriting after deletion, repeated marking of the same expression, or cases where an additional mark is superimposed on an existing highlighted object. The medical management ontology construction module (300) can aggregate the number of modification entries based on chart editing logs or comparison results between consecutive images.

[0141] The medical management ontology construction module (300) can first assign a pre-set basic weight to a pair of a specific skin concern entity and a specific procedure entity. Subsequently, the medical management ontology construction module (300) can sequentially collect the number of overlap occurrences, the emphasis marking density, the number of echo utterances, the synchronization delay time, and the number of modification entries, and correct the basic weight step by step according to each judgment criterion.

[0142] Specifically, the medical management ontology construction module (300) can determine that the voice description and chart display are closely linked within the same consultation flow when the synchronization delay time is within a preset first threshold value, and increase the basic weight by the first increment range. For example, if the first threshold value is set to 10 seconds, a first increment can be performed if the synchronization delay time is 10 seconds or less.

[0143] Additionally, the medical management ontology construction module (300) may further correct the weight by determining that the procedure is repeatedly highlighted in both voice and chart when the number of overlapping occurrences is greater than or equal to a preset second standard value and simultaneously the emphasis marking density is greater than or equal to a preset third standard value. In this case, the medical management ontology construction module (300) may perform a synergy correction greater than the correction range when a single item increases. For example, if the number of overlapping occurrences is 3 or more and the emphasis marking density is 0.5 or more, the weight may be set to increase further to 1.5 times the simple addition range.

[0144] The number of echo utterances mentioned above can be used as a condition for further increasing the relationship weight. That is, if a customer repeatedly mentions the same procedure name or the same efficacy expression immediately after the medical staff's explanation, the medical management ontology construction module (300) determines that the customer is actively accepting or reviewing information related to the procedure and can add the weight further. The increment range based on the number of echo utterances can be configured to increase stepwise as the number of times increases.

[0145] On the other hand, if the synchronization delay time exceeds the first threshold value, it is determined that the connectivity between the voice description and the chart display is relatively low, and the weight can be reduced. In addition, if the number of correction entries is greater than or equal to the pre-set fourth threshold value, it is determined that repeated corrections have occurred for the same chart area, and the relationship strength can be further lowered. For example, if the number of correction entries is two or more, the reduction range of the weight can be set larger than the first reduction range.

[0146] The medical management ontology construction module (300) may sequentially perform a basic weight assignment step, a positive correction step, a negative correction step, and a range limiting step when both positive and negative corrections exist. For example, the medical management ontology construction module (300) may first assign a basic weight, then perform a positive correction based on the synchronization delay time, the number of overlap occurrences, the emphasis marking density, and the number of echo utterances, and subsequently perform a negative correction based on whether the synchronization delay time is exceeded and the number of correction entries. Finally, the medical management ontology construction module (300) may perform range limiting control so that the correction result does not exceed the preset upper and lower limit ranges.

[0147] The above range limit control may be a process designed to mitigate the rapid bias of the relationship weights even when specific events are temporarily excessively concentrated. For example, even when the number of overlap occurrences, the emphasis marking density, and the number of echo utterances are all detected at high levels, the relationship weights may be controlled to converge in stages so as not to exceed a preset maximum threshold. Accordingly, the phenomenon in which specific edges are excessively emphasized by extreme input values ​​can be reduced.

[0148] The medical management ontology construction module (300) may further include an exception handling flow. For example, if medical staff have prepared a template-type chart in advance before the customer enters the room, the time of displaying the chart may precede the time of the voice mention. In this case, the synchronization delay time may be calculated as a value less than 0. If the synchronization delay time is calculated as a negative value, the medical management ontology construction module (300) may determine this as a pre-written template or a pre-consultation preparation record and exclude the chart display event from the synergy correction target. By doing so, the situation in which the relationship weight is calculated higher than the actual consultation response due to the pre-written chart can be reduced.

[0149] The dynamic relationship weights calculated through the above process can be stored as attribute values ​​of knowledge graph edges connecting the skin concern entity and the treatment entity. These attribute values ​​can be used as the degree of priority review of the edge, whether it is a subject for subsequent analysis, or as input characteristic values ​​transmitted to the customized prediction module (500). For example, pairs of entities with relatively high dynamic relationship weights can be processed as high-priority relationships in the subsequent recommendation candidate review stage, and pairs of entities with relatively low dynamic relationship weights can be classified as subjects for additional review or lower-priority processing.

[0150] An example of the operation of the above logic is as follows. In the first embodiment, when the synchronization delay time is 3 seconds, the number of overlap occurrences is 4, the emphasis marking density is 0.8, the number of echo utterances is 2, and the number of correction entries is 0, the medical management ontology construction module (300) can determine that the synchronization delay time, the number of overlap occurrences, the emphasis marking density, and the number of echo utterances satisfy positive correction conditions and adjust the corrected weight to a higher level. At this time, the emphasis marking density of 0.8 may correspond to the case where 4 actual emphasized objects are detected based on 5 standard emphasized objects per 100 normalized areas. In the second embodiment, when the synchronization delay time is 15 seconds, the number of overlap occurrences is 2, and the emphasis marking density is 0.2, the medical management ontology construction module (300) can determine that the degree of satisfaction of the positive correction conditions is low and maintain the corrected weight at a basic level or a lower level. At this time, the above-mentioned emphasis marking density of 0.2 may correspond to the case where one actual emphasized object is detected under the same criteria. In the third embodiment, if the synchronization delay time is calculated as a negative number, the medical management ontology construction module (300) may determine the chart event as a pre-entry template and exclude it from the synergy correction target. In the fourth embodiment, even if the synchronization delay time, the number of overlap occurrences, and the emphasis marking density are all satisfactory, if the number of correction entries accumulates to 3 or more times, a reduction logic may be additionally applied so that the final weight is adjusted to an intermediate level. In the fifth embodiment, even if the number of overlap occurrences and the emphasis marking density fall slightly short of the reference value, if the number of echo utterances is detected multiple times, it may be determined that there is active re-mention by the customer, and the weight may be further increased.In the sixth embodiment, even if the positive correction factor is detected to be very high, the range limit control is applied together so that the final weight can be maintained within the maximum allowable range.

[0151] The existing simple keyword detection method could highly evaluate the relationship between the skin concern and the procedure simply because a specific procedure name appeared repeatedly in voice transcriptions or consultation records. However, the medical management ontology construction module (300) can distinguish and process the simple repeated appearance of the same word and the strength of the connection in the actual consultation context by reflecting the emphasis marking density, the synchronization delay time, the number of echo utterances, and the number of correction entries together. Accordingly, it can be used to distinguish between relationship candidates to be prioritized for subsequent prediction processing and relationship candidates requiring further review.

[0152] The first reference value, the second reference value, the third reference value, and the fourth reference value may not be limited to fixed numerical values. For example, the first reference value for the synchronization delay time may be set to 10 seconds during the initial operation of the system, but may be periodically updated based on the upper percentile value of the voice-chart response interval derived from a set of normal consultation sessions accumulated over the past month. Additionally, the reference values ​​for the number of overlap occurrences and the emphasis marking density may be dynamically reset based on the average, median, or percentile value calculated based on a set of consultation sessions where subsequent procedure linkage was confirmed or a set of consultation sessions that ended without data omission. Accordingly, the medical management ontology construction module (300) can adjust the judgment criteria to reflect differences in consultation environments, record practices, and chart creation methods for each medical institution.

[0153] In order to set the synchronization threshold and overlap occurrence criteria of the medical management ontology construction module (300) above, a basic simulation may be performed on 10 consultation cases in which both voice data and pen chart data are secured in the same consultation session. In the basic simulation above, the first case is one in which the synchronization delay time is 2 seconds and the number of overlap occurrences is 4, and actual procedure conversion is confirmed, and it may be determined that it is appropriate to reflect the relationship weight upward. The second case is one in which the synchronization delay time is 5 seconds and the number of overlap occurrences is 3, and actual procedure conversion is confirmed, and a tendency to reflect the relationship weight upward may be observed. The third case is one in which the synchronization delay time is 9 seconds and the number of overlap occurrences is 3, and actual procedure conversion is confirmed, but since the delay time is close to the boundary range, it may be appropriate to reflect the relationship weight at a level above the middle.

[0154] On the other hand, in the 4th case, where the synchronization delay time is 12 seconds and the number of overlapping occurrences is 2, actual procedure switching may not be confirmed, and in the 5th case, where the synchronization delay time is 18 seconds and the number of overlapping occurrences is 2, switching failure may be confirmed. Additionally, in the 6th case, where the synchronization delay time is 4 seconds and the number of overlapping occurrences is 5, actual procedure switching may be confirmed, and in the 7th case, even if the synchronization delay time is relatively short at 8 seconds, switching may not occur if the number of overlapping occurrences is only 1. In the 8th case, where the synchronization delay time has increased significantly to 25 seconds, actual procedure switching may not occur even if the number of overlapping occurrences is 4. In the 9th case, where the synchronization delay time is 1 second and the number of overlapping occurrences is 2, switching may be confirmed, and in the 10th case, where the synchronization delay time is 11 seconds and the number of overlapping occurrences is 3, it may be classified as a case where the switching status is located in the boundary range.

[0155] From the above 10 basic simulations, if the synchronization delay time generally exists within 10 seconds and the number of overlapping occurrences is 3 or more, a relatively high tendency to be linked to actual procedure switching may be observed; if the synchronization delay time exceeds 10 seconds, a tendency for the relationship strength to weaken may be confirmed even if the number of overlapping occurrences exists at a certain level. In addition, even if the synchronization delay time is short, if the number of overlapping occurrences is low at the level of 1, cases may be observed where it is difficult to evaluate the relationship highly using only a single variable. Accordingly, the medical management ontology construction module (300) may set a first reference value for the synchronization delay time to a range of around 10 seconds and a second reference value for the number of overlapping occurrences to a range of around 3 times.

[0156] The medical management ontology construction module (300) can perform stepwise correction rules based on trends observed in the basic simulation. For example, when the number of overlap occurrences is greater than or equal to the second reference value while the basic weight is assigned, the relationship weight can be adjusted upward in the first step, and when the emphasis marking density is also greater than the reference value, synergy correction can be performed more significantly than when a single condition is satisfied. On the other hand, if the synchronization delay time exceeds the first reference value, a reduction logic for the relationship weight can be performed, and the reduction range can be expanded stepwise as the degree of exceeding the reference value increases. In addition, if the number of modification entries occurs repeatedly, the relationship weight can be further adjusted downward.

[0157] The effectiveness of the above logic can be verified using a separate verification set. For example, regarding consultation data of the same type, the existing simple keyword matching method extracted a total of 152 relationship candidates, but 68 of them could be classified as spurious relationships that are not connected to actual conversions. On the other hand, when the dynamic relationship weighting logic according to the present invention is applied, a total of 94 relationship candidates are extracted, and the number of spurious relationships among them can be reduced to 8. In addition, when comparing the suitability of relationship candidates based on whether there is a subsequent procedure conversion, the prediction agreement rate of the existing method is observed to be 55.2 percent, whereas when the logic of the present invention is applied, a trend of improvement to 89.4 percent can be confirmed. Accordingly, the medical management ontology construction module (300) can be utilized to more selectively manage relationship candidates to be used for subsequent prediction processing by reflecting the degree of synchronization and overlap between voice and chart.

[0158] A data lifecycle flow management module (400) according to one embodiment of the present invention tracks how consultation voice data, pen chart image data, freehand text, and normalized data generated therefrom are processed through a path from the collection stage to the visualization stage, and can determine a flow preservation state indicating to what extent the data maintains its original source and context while passing through each processing stage. The data lifecycle flow management module (400) can calculate the flow preservation state by not merely checking whether the data has been input or displayed on the final screen, but also reflecting whether the data maintained the same source chain at each stage, whether it was disconnected at a specific stage, or whether it was automatically restored without separate intervention by an operator.

[0159] The data lifecycle flow management module (400) may collect a plurality of judgment criteria to determine the flow preservation status. The judgment criteria may include the number of valid nodes in which the original source identifier at the time of initial creation is maintained until the final stage, the source chain completeness ratio representing the ratio of processing stages that are actually connected without interruption among the collection stage, text conversion stage, normalization mapping stage, ontology linkage stage, prediction reflection stage, and visualization reflection stage, the number of occurrences of orphan records that are not properly mapped to customer identification information or consultation session identification information, the cumulative value of timestamp discontinuities calculated by accumulating only the intervals exceeding an allowable threshold among the timestamp differences between adjacent processing stages, and the number of successful automatic recoverys and the number of manual intervention corrections by an operator during the data merging or restoration process.

[0160] The above original source identifier may consist of an original voice file identifier, an original image file identifier, a consultation session identifier, a customer identifier, a conversion task identifier, or a combination thereof. The above data lifecycle flow management module (400) may maintain the above original source identifier as a common reference key among the data collection module (100), the unstructured data normalization module (200), the medical management ontology construction module (300), the customized prediction module (500), and the visualization linkage module (600). Accordingly, the above data lifecycle flow management module (400) can track at which stage specific data was generated, at which stage it was converted, and at which stage it was omitted or merged based on the same source.

[0161] The number of valid nodes mentioned above can be calculated as the number of nodes among the data nodes that have reached the final visualization reflection stage or the subsequent prediction reflection stage in which the original source identifier is maintained without interruption. For example, if voice data, image data, transcription, normalization records, ontology entity connection results, and prediction input records generated in the same consultation session are all linked by the same original source identifier, the data lifecycle flow management module (400) can determine that the node is a valid node.

[0162] The source chain completeness ratio can be calculated by dividing the number of processing steps actually connected to each data node by the total number of processing steps. For example, if the total number of processing steps is 6 and a specific data node is connected up to the collection, text transformation, normalization mapping, ontology linkage, and visualization reflection steps but does not go through the prediction reflection step, the source chain completeness ratio can be calculated as 5 / 6. The data lifecycle flow management module (400) can determine through the ratio how far a specific data has been maintained in the chain within the entire pipeline.

[0163] The above orphan record may refer to a record in which the mapping with customer identification information, consultation session identification information, or original source identifier is severed. For example, if a voice transcription exists but the customer identification information corresponding to the transcription is missing, making it impossible to link the transcription to a specific consultation case, the data lifecycle flow management module (400) may determine this as an orphan record. Additionally, even if procedure information extracted from a pen chart image exists, if the consultation session information linked to the image is missing, it may be treated as an orphan record.

[0164] The above discontinuous timestamp accumulation value can be calculated by accumulating the excess portion of the interval that exceeds a preset allowable threshold among the time differences between adjacent processing steps. For example, if a delay exceeding the allowable threshold occurs in any one of the internal handoff interval registered in the text conversion task queue during the collection stage, the internal handoff interval registered in the normalization mapping task queue after text conversion is completed, or the internal handoff interval registered in the ontology linkage task queue after normalization mapping is completed, the data lifecycle flow management module (400) can reflect the excess portion in the above discontinuous timestamp accumulation value.

[0165] Specifically, if the allowable threshold is set to 150 milliseconds and the time difference from the time of completion of normalization mapping to the time of registration of the ontology linkage task queue is recorded as 300 milliseconds, the data lifecycle flow management module (400) may reflect the 150 milliseconds exceeding the allowable threshold, rather than the total 300 milliseconds, in the timestamp discontinuous accumulation value. In this case, conversion execution sections that require a certain amount of time for the task itself, such as speech recognition processing or optical character recognition processing, may be managed as a separate task time indicator, and the 150 millisecond allowable threshold may be applied as a criterion for evaluating internal transmission delay or queue registration delay between each processing step. Accordingly, the flow preservation state can be determined by distinguishing between normal conversion operation time and connection delay between processing steps.

[0166] The above number of successful automatic recoverys may refer to the number of times the system recovers the data to a normal chained state according to pre-configured automatic recovery rules when a data merge conflict, partial omission of identifiers, or inter-step connection failure occurs. The above automatic recovery rules may include at least one of the following: a first recovery rule that compares the file identifier of the most recently collected log when only the original file identifier is missing while the same customer identification information and the same consultation session identification information are maintained; a second recovery rule that matches candidate consultation sessions when the customer identification information is missing but the consultation room identification information, the consultation manager identification information, and the proximity timestamp match; a third recovery rule that groups them as candidates for the same consultation session when the same procedure name and the same body part code exist in common in the voice transcription and the pen chart record; and a fourth recovery rule that reconnects the original file and the converted result using a file name rule or a conversion task identifier.

[0167] The data lifecycle flow management module (400) can generate candidate connection relationships according to the automatic recovery rule and classify the trust status of the candidate connection relationships as normal, requiring inspection, or recovery failure. For example, if the customer identification information, consultation session identification information, and original file identifier all match, it can be classified as normal recovery, and if the customer identification information is missing but only the consultation manager identification information and proximity timestamp match, it can be classified as requiring inspection. The number of times it is classified as normal recovery can be reflected in the number of automatic recovery successes, and the number of times it is manually confirmed by an operator in the state requiring inspection can be reflected in the number of manual intervention corrections.

[0168] The data lifecycle flow management module (400) can set a default preservation state for a specific data set or a specific data node. Subsequently, the data lifecycle flow management module (400) can correct the default preservation state step by step based on the number of valid nodes, the source chain integrity ratio, the number of connection processing steps, the timestamp discontinuous cumulative value, and the automatic recovery result.

[0169] Specifically, the data lifecycle flow management module (400) can first determine the number of processing steps actually connected among the collection step, text conversion step, normalization mapping step, ontology linkage step, prediction reflection step, and visualization reflection step for each data node. Next, if the number of connected processing steps is greater than or equal to a preset connection step threshold, the data node can be classified as a candidate for an intermediate or higher preservation state. Subsequently, if the number of valid nodes in which the original source identifier is maintained until the final stage is greater than or equal to a preset valid node threshold, the preservation state of the entire data set can be upgraded to a grade higher than the intermediate or higher preservation state. Additionally, if the source chain integrity ratio is greater than or equal to a preset chain integrity threshold, the upgrade range can be further expanded.

[0170] On the other hand, the data lifecycle flow management module (400) may lower the initially raised retention status when an orphan record is detected or when the timestamp discontinuous accumulation value exceeds a preset discontinuous accumulation threshold value. If the number of automatic recovery successes is greater than the number of manual intervention corrections, the range of the lowering may be limited, and if the number of manual intervention corrections is greater than the number of automatic recovery successes, the range of the lowering may be maintained or expanded. The lowering may be a process that reflects the fact that step delays, connection failures, or source omissions may cause a significant flow degradation during the actual operation phase.

[0171] The above number of successful automatic recoverys and the above number of manual intervention corrections can be used as conditions to correct the downward adjustment result again. For example, if the number of successful automatic recoverys is relatively higher than the number of manual intervention corrections, the data lifecycle flow management module (400) may limit the downward adjustment range or offset a portion of the downward adjustment. Conversely, if the number of manual intervention corrections is relatively higher than the number of successful automatic recoverys, the data set is determined to be in a state where it is difficult to recover stably using only internal system rules, and the downward adjustment range may be maintained or expanded.

[0172] The data lifecycle flow management module (400) may ultimately assign one of a plurality of retention status grades to the data set or the data node. For example, the retention status may be divided into multiple stages such as excellent, good, average, caution, and reprocessing required. Data sets classified as excellent or good grades may be prioritized for input into subsequent prediction processing, while data sets classified as caution or reprocessing required grades may be provided with a separate indication on a visualization screen or moved to an automatic reprocessing queue.

[0173] The data lifecycle flow management module (400) may further include an exception handling flow. For example, if a temporary network delay, a storage device response delay, or a surge in batch traffic during a specific time period occurs in the entire hospital system, and timestamp delays are observed simultaneously in multiple data sets, the data lifecycle flow management module (400) may determine this as a system-wide event rather than an anomaly in an individual data flow. In this case, the data lifecycle flow management module (400) may temporarily suspend the downward adjustment logic for the discontinuous cumulative timestamp value or temporarily relax and apply the discontinuous cumulative threshold value. Here, the discontinuous cumulative threshold value may be a threshold value for determining whether to lower the data flow preservation status based on the accumulated value of the excess amount exceeding the allowable threshold among the time differences between adjacent processing steps. By doing so, batch delays caused by system-wide events and source chain disconnections of individual data sets can be distinguished and processed, thereby reducing situations where the entire data set is collectively downvalued due to external factors.

[0174] The flow preservation status determined through the above process may not be used merely as simple monitoring information. The flow preservation status may be used as a criterion for determining the input data priority setting of the customized prediction module (500), whether to display a warning in the visualization linkage module (600), whether to automatically allocate a reprocessing queue, or whether to request an operator inspection. For example, a dataset with a relatively high preservation status may be prioritized for delivery to the customized prediction module (500) to manage the input quality of subsequent prediction processing. On the other hand, a dataset with a relatively low preservation status may be presented with a reliability warning when displayed on the dashboard, or may be separately classified as an operator inspection item.

[0175] Examples of the operation of the above logic are as follows. In the first embodiment, if the number of valid nodes is above a standard, the source chain integrity ratio is 90 percent or higher, and there are no orphan records, the data lifecycle flow management module (400) may classify the data set as excellent or good grade. In the second embodiment, if the number of valid nodes and the source chain integrity ratio are good but one orphan record occurs, the data lifecycle flow management module (400) may lower the retention status by one level. In the third embodiment, if multiple intervals exceeding the allowable threshold during the delay between adjacent stages accumulate, and the accumulated value of the timestamp discontinuity exceeds a standard, the data lifecycle flow management module (400) may lower the retention status to a caution grade or lower. In the fourth embodiment, even if the orphan record and the delay accumulation exist together, if the number of successful automatic recoverys is higher than the number of manual intervention corrections, the data lifecycle flow management module (400) may limit the downward range to maintain the preservation status at a normal grade. In the fifth embodiment, if the delay accumulation is not large but the number of manual intervention corrections occurs repeatedly, the data lifecycle flow management module (400) may determine that the automatic processing stability of the data set is low and classify the preservation status downward. In the sixth embodiment, if a system delay across a specific time period is detected, the data lifecycle flow management module (400) may determine the delay as a system-wide event and defer the downward adjustment.

[0176] Existing simple transformation and loading recording methods may have been limited to storing the success or failure of individual steps separately. In this case, it may be difficult to continuously determine whether specific data actually reached the final analysis or visualization stage while maintaining the source chain. On the other hand, the data lifecycle flow management module (400) can manage the flow status of unstructured data by distinguishing it from various angles by reflecting together the maintenance of the original source identifier, the coverage of the connection processing stage, the occurrence of orphan records, the discontinuous cumulative value of timestamps, and the relative proportion of automatic recovery and manual intervention. Accordingly, this can be utilized to reduce situations where data whose meaning has been altered or whose source has been severed in the middle of the pipeline is fed directly into subsequent decision-making stages.

[0177] The aforementioned connection step threshold, valid node threshold, chain completeness threshold, and discontinuous accumulation threshold may not be limited to fixed numerical values. For example, the connection step threshold may be set based on the minimum number of steps that must actually be connected among the collection step, text transformation step, normalization mapping step, ontology linkage step, prediction reflection step, and visualization reflection step, and may be periodically reset based on the processing step coverage of a set of consultation sessions that were terminated without data omission. The valid node threshold may be set based on the number of nodes that retained original source identifiers among data nodes that reached the final prediction reflection step or visualization reflection step, and may be adjusted based on the average, median, or percentile value calculated from a set of sessions that were successfully reflected up to the subsequent prediction step. The chain completeness threshold may be set based on the ratio of the actual number of connection processing steps to the total number of processing steps, and may be reset by referring to the source chain retention rate of a set of sessions that completed up to visualization reflection. In addition, the above discontinuous cumulative threshold value serves as a criterion for determining whether timestamps are discontinuous. While it may be set as a fixed delay allowance value at the time of initial system introduction, it can be periodically updated thereafter by reflecting the server response time within the hospital, weekly traffic patterns, average data merging time, and accumulated logs of normal processing sessions. Accordingly, the data lifecycle flow management module (400) can adjust the criteria for determining the flow preservation status by reflecting differences in the system environment, log collection method, and operational load pattern for each medical institution. To set the criteria for determining the allowable delay and flow interruption between stages of the data lifecycle flow management module (400), 10 basic simulations can be performed on the internal pipeline processing flow.In the first case, the average delay between adjacent steps is measured at 45 milliseconds, the active node retention rate is 100 percent, and no orphan records occur, so it can be determined as a normal flow. In the second case as well, the average delay is 80 milliseconds, the active node retention rate is 100 percent, and no orphan records occur, so it can be classified as a normal flow. In the third case, the average delay is 120 milliseconds, the active node retention rate is 98 percent, and there are no orphan records, so even if it approaches the boundary section, it can be determined as a flow within the normal range.

[0178] In the fourth case, the average delay increases to 160 milliseconds, the effective node retention rate decreases to 85 percent, and one orphan record may occur, so the data lifecycle flow management module (400) may determine this as a disconnection warning state. In the fifth case, the average delay increases to 310 milliseconds, the effective node retention rate decreases to 60 percent, and four orphan records occur, so the flow may be classified as a severe disconnection state. In the sixth case, the average delay is 55 milliseconds, the effective node retention rate is 100 percent, and there are no orphan records, so it may be maintained as a normal flow. In the seventh case, the average delay is 145 milliseconds, the effective node retention rate is 92 percent, and one orphan record occurs, but it may be classified as a correction case where an automatic recovery rule is applied to avoid flow disconnection. In the 8th case, the average latency is 220 milliseconds, the active node retention rate is 75 percent, and there are 3 orphan records, so it can be determined to be in a disconnected state. In the 9th case, the average latency is 40 milliseconds, the active node retention rate is 100 percent, and there are no orphan records, so it can be classified as a normal flow, and in the 10th case, the average latency is 190 milliseconds, the active node retention rate is 82 percent, and there are 2 orphan records, so it can be determined to be in a disconnected warning state.

[0179] From the above 10 basic simulations, it can be observed that when the average delay between adjacent steps is generally 150 milliseconds or less and no orphan records occur, the flow preservation state tends to remain normal. On the other hand, in the interval where the average delay begins to exceed 150 milliseconds, the effective node retention rate tends to decrease, and when one or more orphan records occur, cases of moving to a flow disconnection warning or disconnection state may be observed. Accordingly, the data lifecycle flow management module (400) can set an allowable threshold for determining whether timestamps are discontinuous to a range of around 150 milliseconds.

[0180] The data lifecycle flow management module (400) can perform state determination rules based on trends observed in the basic simulation. For example, if the ratio of the number of successful connection steps to the total number of processing steps is high and the number of valid nodes where the original source identifier is maintained until the final step is above a standard, the flow preservation state can be classified upward. In this case, if both the number of valid nodes and the ratio of source chain integrity are above a standard, the preservation state can be adjusted upward by a larger margin than when a single condition is met. On the other hand, if delay intervals between adjacent steps exceeding the allowable threshold accumulate or if an orphan record occurs, the flow preservation state can be adjusted downward. Additionally, if the number of successful automatic recoverys is relatively higher than the number of manual interventions, the downward adjustment range can be limited, and if the number of manual interventions is higher, the downward adjustment range can be maintained or expanded.

[0181] The effectiveness of the above logic can be verified using a separate verification set. For example, in the existing simple transformation and loading record method, a data loss rate of 18.5 percent is observed, and the rate at which the original source is sequentially maintained up to the final stage is low, which may result in a high frequency of manual correction by the operator. On the other hand, when the data lifecycle flow management module (400) according to the present invention is applied, the data loss rate is reduced to 2.1 percent, and the rate at which the source identifier and processing stage are sequentially maintained is high, which may result in a lower frequency of manual correction by the operator. Accordingly, the data lifecycle flow management module (400) can be utilized to reduce situations where data with a severed or missing source in the middle of the pipeline is directly fed into subsequent prediction or visualization stages.

[0182] A customized prediction module (500) according to one embodiment of the present invention can determine an acceptance resistance state indicating the customer's potential burden and tendency to reject acceptance of said proposal before providing a Next Best Offer (NBO) to the customer. The customized prediction module (500) can classify said acceptance resistance state in stages by not merely using static information such as past payment history, visit cycle, or age group, but also reflecting changes in the customer's voice response occurring immediately after said proposal, whether inquiries regarding burden appear, whether there is voluntary re-mention, whether there is a follow-up inquiry, and the deviation between the customer's unique re-visit cycle and the current proposal time.

[0183] The above-described customized prediction module (500) may collect a plurality of judgment criteria to determine the acceptance resistance state. The judgment criteria may include the magnitude of voice energy drop observed in the customer response section immediately after the proposal when the average voice energy of the customer response section immediately before the medical staff proposes a specific procedure or efficacy is set as a reference value, the cumulative time of a low-voice response section or a completely unresponsive section maintained below a certain voice energy standard immediately after the proposal, the number of occurrences of inquiries related to burdens such as recovery period, pain, bruising, side effects, visible signs, and management burden, the number of times the customer voluntarily re-mentioned the name of the proposed procedure or the same efficacy expression within the same consultation session, the number of voluntary follow-up inquiries that occurred after the proposal and before the reservation was confirmed, and a phase deviation value indicating the degree of deviation at the current proposal time compared to the proposal possible time period calculated based on the customer's past re-visit interval history.

[0184] The above voice energy drop range can be calculated based on the difference in customer voice energy between the section immediately before the proposal and the section immediately after the proposal. The above customized prediction module (500) can distinguish between customer speech sections and medical staff speech sections by performing speaker separation processing on voice data collected from a wearable badge unit. Subsequently, the above customized prediction module (500) can identify the average voice energy of the customer response section immediately before the medical staff's proposal speech as a reference value, and calculate the average voice energy of the customer response section immediately after the proposal to calculate the drop range relative to the reference value. The above voice energy may be calculated as an absolute decibel value, or processed as a normalized energy value or a relative rate of change.

[0185] The accumulated time of the aforementioned low-voice response interval or no-response interval can be used as a criterion indicating the passivity of the customer's response immediately after the proposal. The customized prediction module (500) can identify not only the interval where customer utterance is completely absent immediately after the proposal, but also the low-voice response interval maintained below a preset voice energy standard, and calculate the accumulated time. For example, if the customer does not respond immediately after the proposal or continues to provide only short, short responses of very low volume, the customized prediction module (500) can determine this as a low-voice response state or a no-response state.

[0186] The number of questions regarding the burden mentioned above can be calculated based on voice recognition results or consultation records. The customized prediction module (500) can extract expressions that may act as a burden in accepting the proposal, such as recovery period, pain, bruising, swelling, side effects, maintenance period, visible signs, time to return to daily life, and whether additional care is needed, through a natural language processing procedure, and can count the number of occurrences of said expressions. The said expressions can be identified based on a predefined dictionary of burden questions or a set of question expressions frequently observed in consultation logs by medical institution.

[0187] The number of voluntary re-mentions mentioned above may represent the number of times a customer repeats the suggested procedure name or the same efficacy expression on their own after the medical staff's suggestion. The customized prediction module (500) can identify whether the same procedure name or the same efficacy expression is repeated within a preset time interval after the suggestion by targeting only the customer's utterance segment after speaker separation processing. For example, if a customer mentions the procedure name again, such as "How long does that procedure last?" or "Is the recovery from the above lifting procedure fast?" after the medical staff suggests a specific lifting procedure, the customized prediction module (500) can count this as a voluntary re-mention.

[0188] The number of voluntary follow-up inquiries mentioned above may be the number of additional contacts actively generated by the customer from the time of the proposal until the reservation is confirmed. The additional contacts may include text inquiries, consultation channel messages, callback requests, additional consultation reservations, or online inquiry inputs. The customized prediction module (500) may compare the time of occurrence of the follow-up inquiries with the time of the proposal to count only customer-driven inquiries that occurred within a preset period after the proposal as the number of follow-up inquiries.

[0189] The above phase deviation value may be a value indicating how much the current proposal time deviates from the proposal possible time interval calculated based on the past revisit interval history of each customer. The above customized prediction module (500) can calculate the average, median, variance, or allowable interval of revisit intervals for the same or similar treatment groups in the past of a specific customer. Subsequently, the above customized prediction module (500) determines whether the current proposal time is located within the allowable interval, and if it deviates from the allowable interval, it can calculate the degree of deviation as the above phase deviation value. For example, if the typical revisit interval of a specific customer is set to an allowable interval of around 180 days, but the current proposal time is on the 120th day, the above customized prediction module (500) can determine that the current time is ahead of the proposal possible time interval and calculate the above phase deviation value.

[0190] The above-described customized prediction module (500) can prioritize setting a basic acceptance resistance state for a specific customer and a specific proposal. The basic acceptance resistance state can be initialized to, for example, low, normal, or high, and can subsequently be corrected stepwise according to the voice energy drop range, the cumulative time of the low-voice response or no-response interval, the number of burden-related inquiries, the number of voluntary re-mentions, the number of subsequent inquiries, and the phase deviation value.

[0191] Specifically, the customized prediction module (500) can adjust the acceptance resistance state by one level upward by determining that the customer has shown a psychological burden or embarrassment immediately after the proposal if the voice energy drop range is greater than or equal to a preset first threshold value. Additionally, the acceptance resistance state can be further adjusted upward if the accumulated time of the low-voice response or no-response interval is greater than or equal to a preset second threshold value. If both the voice energy drop range and the low-voice response or no-response interval are simultaneously greater than the threshold, the customized prediction module (500) can adjust the acceptance resistance state upward by a larger amount than when a single condition is met.

[0192] The number of inquiries regarding the burden mentioned above can be used as a condition to further increase the acceptance resistance state. For example, if the number of inquiries regarding the burden mentioned above is greater than or equal to a pre-set third threshold value, the customized prediction module (500) can determine that the customer is paying more attention to the recovery burden or potential side effects than to the proposal itself, and thus adjust the acceptance resistance state upward.

[0193] The above phase deviation value can be used to determine the suitability of the proposal timing. If the current proposal timing is located outside the customer-specific proposal timing interval, the custom prediction module (500) can determine that the proposal timing is out of sync with the customer's usual acceptance timing and further adjust the acceptance resistance state upward. Additionally, the upward adjustment range can be expanded as the phase deviation value increases.

[0194] On the other hand, the number of voluntary re-mentions and the number of subsequent inquiries can be used as conditions to lower the acceptance resistance state. For example, if a customer voluntarily re-mentions the same procedure name or the same efficacy expression after the proposal, the customized prediction module (500) can determine that the customer is further reviewing the proposal and lower the raised acceptance resistance state by one level. Additionally, if the number of subsequent inquiries is greater than or equal to a pre-set fourth threshold value, the customized prediction module (500) can determine that the customer is showing active interest in the proposal and further lower the acceptance resistance state. When both the number of voluntary re-mentions and the number of subsequent inquiries are satisfied, the downward adjustment range of the acceptance resistance state can be set to be greater than when a single condition is satisfied.

[0195] The above-described customized prediction module (500) can apply the above-described judgment criteria in stages. For example, it can first perform a first upward judgment based on the voice energy drop range and the low-voice response or no-response interval, and then perform a second upward judgment based on the number of burden-related queries and the phase deviation value. Next, it can determine the final acceptance resistance state by performing a downward correction based on the number of voluntary re-mentions and the number of subsequent inquiries. In this way, the immediate response immediately after the proposal, individual timing suitability, and interest recovery signal can be reflected sequentially.

[0196] The above-described customized prediction module (500) can adjust the processing priority of the next-ranked customized proposal based on the final acceptance resistance status. For example, if the acceptance resistance status is determined to be a relatively low grade, the next-ranked customized proposal can be classified as a priority proposal target or a priority review target. On the other hand, if the acceptance resistance status is determined to be a relatively high grade, the output priority of the next-ranked customized proposal can be lowered or classified as a pending target. In addition, the acceptance resistance status can be visualized on a management dashboard or a consultation support screen and used as reference information for medical staff or counselors to adjust the proposal intensity.

[0197] The custom prediction module (500) may further include an exception handling flow. For example, if background noise outside the consultation room, equipment operation sounds, external call sounds, or simultaneous conversations are temporarily introduced, making it difficult to analyze the customer's normal voice waveform, the analysis results of the voice energy drop range or the low-voice response section may be distorted. In this case, the custom prediction module (500) may first determine whether the background noise level exceeds a preset fifth threshold value. If the background noise level exceeds the fifth threshold value, the custom prediction module (500) may temporarily lower the weight of the upward judgment based on the voice energy drop range and the low-voice response section, or suspend the judgment. Instead, the acceptance resistance state may be determined by giving higher weight to the number of burden-related inquiries, the phase deviation value, the number of voluntary re-mentions, and the number of subsequent inquiries.

[0198] An example of the operation of the above logic is as follows. In the first embodiment, if the voice energy drop range immediately after the proposal is above a standard, the cumulative time of the low-voice response interval is also above a standard, and the phase deviation value is calculated to be outside the customer-specific allowable range, the customized prediction module (500) can adjust the acceptance resistance state upward to a high grade. In the second embodiment, if the voice energy drop range is not large but the number of burden-related inquiries occurs repeatedly and the phase deviation value is calculated to be large, the customized prediction module (500) determines that the proposal timing is inappropriate and can adjust the acceptance resistance state upward. In the third embodiment, even if a temporary burden signal exists in the voice response, if the customer voluntarily re-mentions the procedure name before the consultation ends and leaves a follow-up inquiry after returning home, the customized prediction module (500) determines that an interest recovery signal exists and can adjust the raised acceptance resistance state downward to a normal or low grade. In the fourth embodiment, if the background noise is detected above a threshold, the custom prediction module (500) may limit voice energy-based judgment and determine the acceptance resistance state based on the burden-related query and the phase deviation. In the fifth embodiment, if the phase deviation value is small but the low-voice response interval and the burden-related query are detected simultaneously, the custom prediction module (500) may determine that there is a psychological burden regarding the proposal itself and adjust the acceptance resistance state upward. In the sixth embodiment, if the phase deviation value is within the allowable range and both the voluntary re-mention and the subsequent inquiry occur, the custom prediction module (500) may classify the next-ranked custom proposal as a priority review target.

[0199] Existing static recommendation methods could select recommendation targets based on a customer's past payment history, last visit date, or simply whether a period had elapsed. In this case, it may be difficult to adequately reflect whether the customer is actually ready to accept the proposal, whether the timing of the proposal aligns with the customer's usual acceptance timing, or whether a burden signal appears immediately after the proposal. On the other hand, the customized prediction module (500) can provide a judgment flow that simultaneously considers responsiveness immediately after the proposal and individual timing suitability by reflecting together the voice response immediately after the proposal, the burden-related query, the voluntary re-mention, the subsequent inquiry, and the phase deviation value. Accordingly, it can be utilized to adjust the output priority of the next-ranked customized proposal or to determine whether to hold it.

[0200] The first reference value, the second reference value, the third reference value, the fourth reference value, and the fifth reference value may not be limited to fixed values. For example, the reference time for the low response or no response interval may be set to 3 seconds upon system introduction, but may be periodically updated by reflecting the average conversation speed by customer group, response delay patterns by age group, and logs of normal consultation sessions accumulated thereafter. Additionally, the allowable interval for the phase deviation value may be individually adjusted based on the average value, variance value, or percentile value calculated from the past visit interval history of a specific customer. The reference value for the number of burden-related inquiries or the voice energy drop range may also be reset according to a reference value calculated based on a set of normal proposal sessions, a set of sessions where actual reservation linkage has been confirmed, or a set of sessions where a subsequent inquiry occurred. Accordingly, the customized prediction module (500) may adjust the criteria for determining the acceptance resistance state by reflecting the consultation environment of the medical institution, customer group characteristics, and response differences by procedure group.

[0201] In order to establish criteria for determining the acceptance resistance state of the custom prediction module (500) above, a basic simulation may be performed on 10 consultation cases in which customer responses were secured immediately after the proposal. In the first case, immediately after the proposal, the voice energy drop is at the level of 5 decibels and the phase deviation is at the level of 10 days; the low-voice response interval is short, questions related to burden are not prominent, and voluntary re-mention is confirmed, so it can be classified as a positive acceptance case. In the second case, the voice energy drop is at the level of 12 decibels and the phase deviation is at the level of 25 days; a low-voice response interval is partially observed and questions related to burden occur at the level of 1 to 2 times, so it can be classified as a reserved response case. In the third case, the voice energy drop is at the level of 18 decibels and the phase deviation is at the level of 45 days; the low-voice response or no-response interval is long and questions related to burden are repeated, so it can be classified as a case with a high probability of proposal rejection.

[0202] In the fourth case, the voice energy drop is at the 8-decibel level and the phase deviation is at the 15-day level; since voluntary re-mentioning was confirmed following the proposal, it can be classified as a positive acceptance case. In the fifth case, the voice energy drop is at the 22-decibel level and the phase deviation is at the 60-day level; since numerous burden-related inquiries occurred and no follow-up inquiries were made, it can be classified as a strong rejection case. In the sixth case, the voice energy drop is at the 14-decibel level and the phase deviation is at the 30-day level; since low-voice response intervals and burden inquiries were confirmed together, it can be classified as a case with high acceptance resistance. In the seventh case, the voice energy drop is at the 3-decibel level and the phase deviation is at the 5-day level; since both voluntary re-mentioning and follow-up inquiries were confirmed, it can be determined to be in a very low resistance state. In the eighth case, the voice energy drop is at the 16-decibel level and the phase deviation is at the 55-day level; since burden-related inquiries exist and no follow-up inquiries were confirmed, it can be classified as a case with high resistance. In the 9th case, the voice energy drop is at the 10-decibel level and the phase deviation is at the 20-day level, placing it in the boundary range; however, some spontaneous re-mentioning is confirmed, so it can be classified as a normal level of resistance. In the 10th case, even though the phase deviation is not large at the 10-day level, the voice energy drop is steep at the 25-decibel level and the low-frequency response interval is long, so it can be determined as a high resistance case.

[0203] From the above 10 basic simulations, if the voice energy drop rate generally increases to around 15 decibels or more and the phase deviation increases to 30 days or more, a tendency to be associated with a rejection of proposal acceptance or a reservation response may be observed. On the other hand, if the voice energy drop rate is low and the phase deviation is within the allowable range, and voluntary re-mention or follow-up inquiry is confirmed, a tendency to be relatively high for proposal acceptance may be observed. In addition, since there are cases where a burden response is high even if the phase deviation is small but the voice energy drop rate is very large, the above customized prediction module (500) can determine the acceptance resistance state by considering the voice energy drop rate and the phase deviation value together.

[0204] The customized prediction module (500) can perform stepwise judgment rules based on the trend confirmed in the basic simulation. For example, if the voice energy drop range is greater than or equal to a preset first threshold value, or if the cumulative time of the low-voice response or no-response interval is greater than or equal to a preset second threshold value, the acceptance resistance state can be adjusted upward by one step. If both the voice energy drop range and the low-voice response interval are greater than the threshold, the acceptance resistance state can be adjusted upward by a larger amount than when a single condition is satisfied. Additionally, if the number of burden-related inquiries is greater than or equal to a preset third threshold value, or if the phase deviation value exceeds the customer-specific allowable proposal interval, the acceptance resistance state can be further adjusted upward. Conversely, if the number of voluntary re-mentions or the number of subsequent inquiries is confirmed, the raised acceptance resistance state can be adjusted downward, and if both conditions are satisfied, the downward adjustment range can be set larger.

[0205] The effectiveness of the above logic can be verified in a separate verification set. For example, in the existing simple cycle-based proposal method, a total of 200 proposals were executed, but actual payment conversions were only 24, resulting in a proposal success rate of 12.0 percent. On the other hand, when the acceptance resistance state determination logic according to the present invention is applied, proposals for customer groups with high resistance are partially withheld, and customer groups with low or downwardly corrected resistance are classified as priority review targets, thereby adjusting the total number of proposals to 120. As a result, actual payment conversions increase to 58, and a tendency for the proposal success rate to increase to 48.3 percent can be confirmed. Accordingly, the customized prediction module (500) can be utilized to provide a selective proposal flow that reflects both proposal timing and customer response, rather than indiscriminate recommendations.

[0206] A customized prediction module (500) according to one embodiment of the present invention can determine the burnout state of medical staff by reflecting the physical and cognitive load accumulated during the medical staff's daily work performance, and a visualization linkage module (600) can control a reservation management screen, an operation dashboard screen, or a rest recommendation screen based on the determination result. The customized prediction module (500) can classify the current state of medical staff into multiple stages by reflecting not only the daily working hours or the number of reservations, but also equipment operation history, spatial movement history, repeated explanation history, claim-related voice events, schedule overrun history, and continuous rest periods. The visualization linkage module (600) can output the classification result in conjunction with the display status of available reservation slots, a rest need notification, or a schedule rescheduling request screen.

[0207] The above-described customized prediction module (500) may collect multiple judgment criteria to determine the exhaustion state. The judgment criteria may include the equipment load status calculated by reflecting the total number of equipment operations performed at the same time and the history of distributed operations in multiple procedure rooms or examination rooms, the number of space switching times indicating the number of moves between examination rooms, procedure rooms, consultation rooms, or recovery rooms on the day, the accumulated time of repeated explanations indicating the time when the same or similar explanations are repeated due to the customer's lack of understanding, re-questioning, or request for supplementary explanation, the number of occurrences or duration of claim-related voice peaks that rise above a preset voice volume standard during the consultation process, the number of schedule overruns or the accumulated time of schedule overruns that exceed the scheduled treatment or procedure end time and squeeze the next slot, and the length of uninterrupted rest blocks continuously secured without patient identification tags on the reservation schedule.

[0208] The above equipment load status may not be calculated solely by the total number of shots. The customized prediction module (500) can calculate the equipment load status by collecting operation logs of skin beauty equipment, lifting equipment, laser equipment, or other procedure equipment, and by reflecting the history of distributed execution in multiple equipment or multiple procedure rooms at the same time. For example, a case where a total of 2,000 shots are performed intensively in one procedure room within the same 30-minute interval, and a case where the same total number of shots are performed distributedly in three procedure rooms with continuous movement, can be distinguished as different equipment load states. The customized prediction module (500) can calculate the equipment load status by reflecting the total number of operations, the number of distributed executions, and the execution density time together.

[0209] The number of space changes mentioned above can be calculated as the number of times medical staff moved between the examination room, procedure room, consultation room, recovery room, or equipment preparation area on the day. The movement history mentioned above can be collected based on entry / exit logs, location tag logs, electronic chart access location logs, or room assignment history from a schedule management system. The accumulated time for repeated explanations mentioned above can be calculated as the sum of the times when the same or similar explanation sections were repeated. The customized prediction module (500) can identify sections where the same procedure explanation, the same precaution explanation, or the same side effect guidance was repeated using voice recognition result text, consultation logs, or explanation script matching results, and can aggregate the accumulated time of the repeated explanations.

[0210] The above claim-related voice peak may include a section during a customer or guardian's speech where the voice volume rises above a preset standard, or a section where it is maintained at a high pitch along with a preset negative expression. The custom prediction module (500) can calculate the occurrence, maximum size, duration, and number of repeated occurrences of the claim-related voice peak by analyzing voice waveforms acquired from a wearable badge unit or a consultation room acoustic sensor. The number of schedule overruns can be calculated by aggregating the number of cases where the actual end time exceeds the scheduled end time using an electronic medical record scheduler, a reservation management system, or a procedure end log. Additionally, the custom prediction module (500) can calculate the accumulated time overruns separately from the number of schedule overruns.

[0211] The above uninterrupted rest block may refer to a continuous time interval on the schedule in which a patient identification tag, procedure tag, consultation tag, or call tag is not assigned. The above-described customized prediction module (500) may identify only the continuous interval actually secured by the medical staff without work intervention as the above-described uninterrupted rest block by referring together the reservation schedule, electronic medical record schedule, location log, and equipment log. For example, even if there is a formal gap of 20 minutes, if a call response, equipment check, or brief consultation is inserted during that time, the entire interval may not be determined as an uninterrupted rest block.

[0212] The above-described customized prediction module (500) can prioritize setting a basic burnout state for a specific medical staff or a specific time interval. The basic burnout state can be initialized to, for example, any one of stable, caution, warning, or high warning, and can then be corrected stepwise according to the equipment load state, the number of space switches, the accumulated time of repeated explanations, the claimable voice peak, the number of cases exceeding a schedule, and the uninterrupted rest block.

[0213] Specifically, the customized prediction module (500) can first increase the basic exhaustion state when the equipment load state is greater than or equal to a preset first threshold value. Additionally, the basic exhaustion state can be further increased when the number of space changes is greater than or equal to a preset second threshold value or when the accumulated time of repeated explanation is greater than or equal to a preset third threshold value. In particular, when both the number of space changes and the accumulated time of repeated explanation are greater than the threshold, the basic exhaustion state can be increased by a larger margin than when a single condition is met. This processing may reflect the fact that cognitive exhaustion can accumulate more significantly in situations where medical staff frequently move between multiple spaces while simultaneously repeating the same explanation.

[0214] The above claim-related voice peak and the above number of times exceeding a certain threshold can be used as conditions for early warning transition. For example, if the claim-related voice peak occurs at or above a preset fourth threshold value, or if the above number of times exceeding the threshold or the above accumulated time of exceeding the threshold is at or above a preset fifth threshold value, the custom prediction module (500) may prioritize transitioning the state to an early warning stage or a higher warning stage, even if the basic exhaustion state has not yet reached a warning stage. Additionally, if both the maximum size and duration of the claim-related voice peak appear high, or if multiple instances of exceeding the threshold accumulate in the same time period, the warning stage may be raised by one more step. In this way, it is possible to distinguish and process loads that accumulate gradually and loads that increase rapidly due to specific events.

[0215] The above uninterrupted rest block can be used as a criterion for mitigating the above exhaustion state. The custom prediction module (500) may lower the accumulated exhaustion state or shorten the maintenance time of the warning stage if the length of the above uninterrupted rest block is greater than or equal to a preset sixth criterion value. However, instead of continuously increasing the mitigation range equally as the length of the above uninterrupted rest block increases, it may be configured to limit the additional mitigation range after a preset recovery period. For example, a rest between 10 and 20 minutes may be reflected relatively significantly in mitigating the above exhaustion state, but for a range exceeding 20 minutes, the mitigation range may increase gradually or be limited to a certain level.

[0216] The customized prediction module (500) can apply the above judgment criteria sequentially. For example, it can first calculate a basic exhaustion state based on the equipment load status, the number of space switches, and the accumulated time of repeated explanations. Then, it can determine whether to switch to an early warning based on the claim-related voice peak and the number of cases exceeding a schedule. Next, it can mitigate the maintenance time of the warning stage or the stage value of the exhaustion state based on the existence and duration of the uninterrupted rest block. In this way, the daily accumulated load, momentary weighting events, and recovery intervals can be reflected in chronological order.

[0217] The visualization linkage module (600) receives the exhaustion status or the warning stage from the customized prediction module (500) and can display a reduced number of available reservation slots on a screen linked to the reservation management system, output a request to change the reservation availability status, or output a screen recommending rest for medical staff. For example, if the exhaustion status is at the caution stage, a certain section can be displayed in a warning color, and if the exhaustion status is at the warning stage, the number of new available reservation slots for the same time period can be reduced and displayed. Additionally, if the exhaustion status is at a higher warning stage, a request for schedule reallocation or a request to limit additional assignments can be output on the screen. If the exhaustion status is alleviated by securing the uninterrupted rest block, the visualization linkage module (600) can expand the displayed slots again or release the request to change the reservation availability status.

[0218] The custom prediction module (500) and the visualization linkage module (600) may further include an exception handling flow. For example, if an emergency treatment code, an emergency procedure code, or an equipment error code is recorded in an electronic medical record or equipment log, the schedule overrun may be determined to be due to external factors or unavoidable circumstances rather than general operational inefficiency. In this case, the custom prediction module (500) may temporarily lower the weight of the early warning transition based on the number of schedule overruns, or temporarily exclude the event from the schedule overrun accumulation target. Additionally, the visualization linkage module (600) may display the exception situation with a separate status code and output a screen distinct from general warnings. By doing so, situations where the entire reservation management flow is excessively constrained by external factors can be reduced.

[0219] Examples of the operation of the above logic are as follows. In the first embodiment, if the equipment load status within the same 30-minute interval is high and both the number of space switchings and the accumulated time of repeated explanations are above a standard, the customized prediction module (500) can raise the basic exhaustion state to a stage just before a warning. In the second embodiment, even if the basic exhaustion state remains in the caution stage, if the claim-related voice peak during consultation is detected above a standard, the customized prediction module (500) can switch the state to an early warning stage. In the third embodiment, if the number of cases exceeding the schedule accumulates and the accumulated time exceeding the schedule also exceeds a standard, the customized prediction module (500) raises the exhaustion state to a higher warning stage, and the visualization linkage module (600) can reduce the display of available reservation slots on the reservation management screen. In the fourth embodiment, if an uninterrupted rest block of 20 minutes or more is identified after the burnout state has been switched to a warning stage, the customized prediction module (500) may shorten the duration of the warning stage or alleviate the burnout state by one level, and the visualization linkage module (600) may partially disable the display of reduced reservation slots. In the fifth embodiment, if an emergency treatment code is identified together on the electronic medical record even if the schedule overrun has occurred, the customized prediction module (500) may defer the early warning transition by distinguishing it from a general schedule overrun. In the sixth embodiment, if the equipment load state is not high but the number of space switches and the accumulated time of repeated explanations are simultaneously high, and claim-related voice peaks occur continuously, the customized prediction module (500) may determine that the cognitive load and emotional load are superimposed and switch the burnout state to a higher warning stage.

[0220] The existing fixed-type reservation assignment method could maintain available reservation intervals based only on working hours or the number of basic slots. In this case, it may be difficult to adequately reflect operational conditions such as the daily load status of medical staff, frequency of space movement, time required for repeated explanations, or intensity of customer interaction. On the other hand, the customized prediction module (500) can determine changes in the daily status of medical staff based on multiple criteria by reflecting the equipment load status, the number of space changes, the accumulated time for repeated explanations, the peak of claim-related voices, the number of schedule overruns, and the uninterrupted rest blocks together. Additionally, the visualization linkage module (600) links the judgment results with the display status of available reservation slots or the rest recommendation screen, allowing the operator to utilize them to adjust whether to assign additional slots in specific time slots. Accordingly, this can be used to alleviate periods of concentrated schedule pressure and to make operational decisions that reflect the status of medical staff.

[0221] The above-mentioned first, second, third, fourth, fifth, and sixth reference values ​​may not be limited to fixed numerical values. For example, the criteria for determining the claim-related voice peak may be set based on a certain voice volume standard during the initial introduction of the system, but may be periodically updated thereafter by reflecting the average environmental noise of the hospital's lobby, consultation rooms, and procedure rooms, voice distribution during normal consultations, and actual warning transition history. Additionally, the reference value for the equipment load status may be adjusted based on the past procedure processing history of each medical staff member, the average operating density by equipment type, and reference values ​​calculated from a set of normal sessions. The reference values ​​for the number of space transitions or the accumulated time for repeated explanations may also be updated based on average, median, or percentile values ​​calculated from a set of normal workdays, a set of high-load workdays, or a set of workdays where condition alleviation was confirmed after rest intervention. The recovery interval of the uninterrupted rest block may also be set differently depending on the work patterns of each medical staff member or the work density of each procedure group. Accordingly, the above-mentioned customized prediction module (500) can adjust the above-mentioned burnout status judgment criteria by reflecting the operating environment of each medical institution and the work characteristics of each medical staff.

[0222] In order to set the criteria for predicting the burnout state and controlling the schedule for the above-mentioned customized prediction module (500) and visualization linkage module (600), 10 basic simulations can be performed assuming the daily work flow of medical staff. In the first case, the number of equipment shots is at the level of 500, the number of space changes is at the level of 2, the accumulated time for repeated explanations is short, there are no claim-related voice peaks or schedule overruns, and there is a non-interrupted rest block of 25 minutes, so the actual burnout state may be classified as low and no warning may be triggered. In the second case, the number of equipment shots is at the level of 1,500, the number of space changes is at the level of 5, and although there are some repeated explanations, there are no claim-related voice peaks and schedule overruns are limited, so the burnout state may be classified as normal. In the third case, the number of equipment shots is at the level of 2,500, the number of space changes is at the level of 8, the repeated explanation time is long and schedule overruns accumulate, so the burnout state may be determined as high and a warning may be triggered.

[0223] In the fourth case, even if the number of equipment shots is not high at the level of 1,000, a claim-related voice peak occurs accompanied by repeated explanations, so the actual exhaustion state may be classified as very high, and the customized prediction module (500) may prioritize switching to an early warning stage. In the fifth case, the number of equipment shots is at the level of 3,000 and the number of space switches is at the level of 12, and the repeated explanation time is long, schedule overruns accumulate, and rest blocks are short, so it may be classified as a higher warning stage and a reduction of available slots may be performed. In the sixth case, the number of equipment shots is at the level of 800 and the number of space switches is at the level of 1, there are no claim-related voice peaks or schedule overruns, and an uninterrupted rest block of 30 minutes is secured, so the actual exhaustion state may be classified as very low. In the seventh case, the number of equipment shots is at the level of 2,000 and the number of space switches is at the level of 6, and a claim-related voice peak occurs, so an early warning may be prioritized before the basic exhaustion state is upgraded. In the 8th case, the number of equipment shots is at the 1,200 level and the number of space switches is at the 10 level; although the repetition explanation time is long, there are no claim-related negative peaks, so it can be determined as a borderline stage. In the 9th case, the number of equipment shots is at the 1,800 level and the number of space switches is at the 4 level; since the schedule overrun is not significant, it can be maintained at a normal level. In the 10th case, the number of equipment shots is at the 2,200 level and the number of space switches is at the 7 level; since claim-related negative peaks and schedule overruns occur together, a higher warning stage is applied and a reduction of valid available slots may be performed.

[0224] From the above 10 basic simulations, if the number of equipment shots and the number of space switches increase together, a tendency for the above basic exhaustion state to gradually increase can be observed. Additionally, if a claim-related voice peak exists, cases may be confirmed where the system transitions first to a warning stage higher than the above basic exhaustion state, even if the total equipment load is at an intermediate level. Conversely, if a certain level of uninterrupted rest blocks is secured, a tendency may be observed where the maintenance time of the above exhaustion state is shortened or adjusted downward. Accordingly, the above customized prediction module (500) can set criteria for determining the exhaustion state by reflecting the equipment load status, the number of space switches, the accumulated time of repeated explanations, the claim-related voice peak, the number of cases exceeding a certain limit, and the uninterrupted rest blocks together.

[0225] The customized prediction module (500) can perform state determination rules based on trends observed in the basic simulation. For example, if the equipment load state is greater than or equal to a preset first threshold value and the number of space switches is greater than or equal to a preset second threshold value, the basic exhaustion state can be adjusted upward by a first step. If both the number of space switches and the accumulated time of repeated explanation are greater than the threshold, the basic exhaustion state can be adjusted upward by a larger amount than when a single condition is met. Additionally, if the size or duration of the claim-related voice peak is greater than or equal to a preset third threshold value, or if the number of cases exceeding a certain threshold or the accumulated time exceeding a certain threshold is greater than or equal to a preset fourth threshold value, the basic exhaustion state can be switched to an early warning stage or a higher warning stage even if it has not yet reached a threshold. On the other hand, if the length of the uninterrupted rest block is greater than or equal to a preset fifth threshold value, the exhaustion state can be adjusted downward or the duration of the warning stage can be shortened, and the additional mitigation range can be limited for rests that exceed a preset recovery interval.

[0226] The effectiveness of the above logic can be verified in a separate operational verification section. For example, in the existing fixed-type reservation assignment method, the rate of medical staff leaving or taking leave may be observed at 12.4 percent, the rate of procedure errors at 3.5 percent, and the average customer satisfaction at 3.8 points out of 5. On the other hand, in the operational section where the burnout state prediction and schedule control logic according to the present invention is applied, a tendency may be observed where the rate of medical staff leaving or taking leave decreases to 2.5 percent, the rate of procedure errors decreases to 0.4 percent, and the average customer satisfaction increases to 4.7 points. Accordingly, the customized prediction module (500) and the visualization linkage module (600) can be utilized for operational decisions to adjust the display of available reservation slots and additional assignments by reflecting the daily load status of the medical staff.

[0227] A medical institution-customized management decision support system according to one embodiment of the present invention may include a data collection module (100), an unstructured data normalization module (200), a medical management ontology construction module (300), a data lifecycle flow management module (400), a customized prediction module (500), and a visualization linkage module (600). Each of the above components may be linked to reflect not only data processing for a single consultation case but also the operational status of multiple consultation cases within the same date. Below, the connection relationship between the above components and processes 1 to 4 is explained by using a series of flows, ranging from a specific customer's consultation reservation to the generation of a next-priority customized proposal and the control of medical staff schedules, as an example in a medical institution providing skin beauty treatments.

[0228] The medical institution may include a customer relationship management system, an electronic medical record system, an appointment management system, a sales management system, a wearable badge unit, a handwritten pen chart image input means, and an operation dashboard. In the medical institution, a customer's appointment history, treatment history, payment history, consultation voice, and handwritten pen chart records may originate from different sources, and the present invention may process the heterogeneous data by linking them based on the same customer and the same consultation session.

[0229] For example, a specific customer A may book a consultation for improving frown lines through a social networking service advertisement. The above reservation may be registered for 14:00 on April 22, 2026, and the reservation management system may store the reservation channel, reservation procedure candidates, reservation manager, and reservation change history. On the day of the consultation, the above customer A may proceed with a consultation with medical professional B, and during the consultation process, consultation voice data may be collected by a wearable badge unit, and after the consultation ends, an image of a handwritten pen chart written by the above medical professional B may be captured and uploaded.

[0230] The data collection module (100) can collect the history of customer A’s revisit consultation appointments from the customer relationship management system and can collect customer A’s previous procedure history, visit history, medical memo, payment history, and past re-consultation interval history from the electronic medical record system. Additionally, the data collection module (100) can receive a voice file containing the start and end times of the consultation from the wearable badge unit and can also collect the handwritten pen chart image file uploaded after the consultation ends. In this case, customer A can be set as a customer with a history of visits for the same or similar procedure group in the past, and the customized prediction module (500) can calculate a phase deviation value between the possible proposal time interval and the current proposal time using the past re-visit interval history.

[0231] In another embodiment, if customer A is a first-time consultation customer, there may not be an individual history of past revisit intervals for customer A. In this case, the customized prediction module (500) may set a provisional proposal possible time interval by referring to the average re-consultation cycle, median, or percentile interval of a customer group having the same inflow channel, same age group, same skin concern, or same consultation procedure group. The phase deviation value calculated using the provisional proposal possible time interval may be stored as a cohort-based phase deviation value distinct from the individual history-based phase deviation value, and the customized prediction module (500) may set the reflection weight of the cohort-based phase deviation value to be lower than that of the individual history-based phase deviation value.

[0232] The data collection module (100) can assign customer identification information, consultation session identification information, collection time information, and source identification information to each data during the collection step. For example, identification information such as “Customer A, Consultation Session 20260422-1400, Consultation Manager B” can be associated with the consultation voice file, and the same consultation session identification information can be assigned to the handwritten pen chart image. The data collection module (100) can subsequently perform voice recognition processing on the voice file to generate a voice transcription, and perform optical character recognition processing on the pen chart image to generate a character extraction result.

[0233] For example, the voice transcription may include the expression “I am concerned about my frown lines and I am curious about how long the Botox lasts,” and the optical character recognition result may extract entries such as “frown lines,” “Botox,” “reconsultation,” and “check after months.” The data collection module (100) may store the original file and the converted result in a mutually mapped form.

[0234] The above unstructured data normalization module (200) can receive a voice transcription and an optical character recognition result received from the data collection module (100) and perform text preprocessing. In the preprocessing step, spacing correction, removal of special characters, abbreviation correction, synonym integration, and term standardization may be performed.

[0235] For example, expressions such as “between-eyebrow Botox” written in the above voice transcription can be converted into the standard procedure name “between-eyebrow Botox,” and abbreviated expressions such as “after check” written in the above pen chart can be corrected to “check after months.” Subsequently, the above unstructured data normalization module (200) can extract the procedure name, procedure site, dosage, purpose of consultation, inquiry regarding burden, timing of revisit, and sentiment expression from the above text.

[0236] Specifically, the unstructured data normalization module (200) can extract “frown lines” as a skin concern item, “Botox” as a procedure item, and “how long the maintenance period is” as a burden or information search related query item from the voice transcription. Additionally, from the pen chart, it can extract “frown lines” as a body part item, “Botox” as a procedure item, and “check after months” as a re-consultation timing item. The unstructured data normalization module (200) can merge the multiple items into a single structured record and assign a standard procedure code, a standard body part code, a consultation purpose code, a burden query code, and a re-visit timing code.

[0237] The above medical management ontology construction module (300) can define entities such as Customer A, inflow channel, skin concern, counselor, medical staff, procedure name, area, payment, and revisit time based on the above standardized data. For example, Customer A can be configured as a Customer entity, social networking service advertisement as an inflow channel entity, frown lines as a skin concern entity, and Botox as a procedure entity.

[0238] The medical management ontology construction module (300) can create relationships between the entities. For example, a “inflow” relationship can be created between the customer entity and the inflow channel entity, a “appeal” relationship between the customer entity and the skin concern entity, and a “corresponding procedure candidate” relationship between the skin concern entity and the procedure entity. At this time, dynamic relationship weight processing logic according to Process 1 can be applied together.

[0239] Specifically, the medical management ontology construction module (300) can collect the number of overlapping occurrences, emphasis marking density, synchronization delay time, number of echo speeches, and number of correction entries for the “between eyebrows-Botox” combination that appears commonly in the voice data and the pen chart image. For example, if Botox is first mentioned in the consultation voice at 14:05:10 and a Botox-related marking is first confirmed in the chart editing log or image upload log at 14:05:14, the synchronization delay time can be calculated as 4 seconds. Additionally, if an underline and an asterisk are displayed together in the corresponding procedure area of ​​the pen chart, and the customer mentions the name of the procedure again, such as “Then, how long does the Botox last?” immediately after the medical staff’s explanation, the number of echo speeches can also increase.

[0240] The medical management ontology construction module (300) can increase the relationship weight between the skin concern entity and the procedure entity if the number of overlapping occurrences and the emphasis marking density are above a standard, the synchronization delay time is within a standard, and the echo utterance exists. On the other hand, if there is repeated modification in the same chart area, or if it is determined to be a pre-template state where the chart entry time exists before the consultation voice, the synergy correction target can be limited. As a result, the “glabellar wrinkle-Botox” edge can be stored as a relationship having a relatively high attribute value.

[0241] The data lifecycle flow management module (400) can connect the original voice file, voice transcription, optical character recognition result, structured record, ontology relationship information, prediction input record, and dashboard reflection record included in the consultation case into a single flow chain. The data lifecycle flow management module (400) can assign a creation time, transformation time, storage time, normalization completion time, relationship creation time, prediction reflection time, and visualization reflection time to each stage.

[0242] According to the above process 2, the data lifecycle flow management module (400) can determine the number of valid nodes of the consultation case, the source chain completeness ratio, whether orphan records occur, the cumulative value of timestamp discontinuous, and the number of successful automatic recoverys and manual interventions. For example, in the consultation case of customer A, if the original voice file, voice transcription, optical character recognition result, normalized record, ontology relationship information, prediction input record, and visualization reflection record are all connected by the same source identifier, the delay between adjacent stages is within the allowable range, and no orphan records occur, the data lifecycle flow management module (400) can classify the consultation case into a higher preservation state.

[0243] Conversely, in other consultation cases, a pen chart image may be uploaded but customer identification information may be missing, so the image may not be linked to a specific consultation session. In this case, the data lifecycle flow management module (400) may classify the record as an orphan record and may downgrade the flow preservation status or move it to a reprocessing queue. Accordingly, the data lifecycle flow management module (400) can manage the chain maintenance status of data input to the subsequent prediction stage in advance.

[0244] The customized prediction module (500) can calculate customer segmentation, revisit pattern analysis, and next-ranked customized proposal candidates using relationship information received from the medical management ontology construction module (300) and time-series flow data received from the data lifecycle flow management module (400). For example, if customer A was acquired through a social network service advertisement, repeatedly mentioned frown line improvement during the initial part of the consultation, and the “frown line-Botox” relationship weight is stored as high, the customized prediction module (500) can classify customer A as a candidate group for specific procedure conversion.

[0245] Subsequently, the customized prediction module (500) can determine the resistance state to accepting the next-priority customized proposal of customer A according to process 3. For example, when the medical staff suggests in the latter part of the consultation that “a combination of managing the forehead area along with the glabella Botox is also possible,” the customized prediction module (500) can calculate the drop in energy by comparing the average voice energy of the customer response section immediately before the proposal with the average voice energy of the customer response section immediately after the proposal. In addition, the cumulative time of low-voice response or no-response section immediately after the proposal, whether inquiries regarding burden occurred, the number of times the same procedure name or efficacy expression was voluntarily re-mentioned, and the number of follow-up inquiries before the reservation is confirmed can be collected together.

[0246] For example, if customer A does not respond immediately for more than 2 seconds after the proposal and subsequently repeats burden-related questions such as “How long does the recovery period take?”, but asks again just before the consultation ends, “Then, if I do the forehead area as well, does it help with maintenance?”, and leaves an additional inquiry via text message on the same day after returning home, the custom prediction module (500) may initially adjust the acceptance resistance state upward, but may correct the acceptance resistance state downward again by reflecting the voluntary re-mention and subsequent inquiry.

[0247] Additionally, the customized prediction module (500) can determine whether the current proposal time is located within the customer-specific allowable proposal range by referring to the past re-visit interval history of customer A. For example, if the average re-consultation cycle of the similar procedure group for customer A is in the range of 90 to 120 days and the current consultation time is close to the above range, the phase deviation can be calculated as small. In this case, the customized prediction module (500) can maintain the acceptance resistance state at a lower level. On the other hand, if the proposal time is excessively ahead or behind the customer's unique re-visit cycle, the acceptance resistance state can be maintained at a high level.

[0248] As a result, the customized prediction module (500) can assign a status to customer A such as “NBO for priority review” or “NBO for hold,” and the recommendation criteria can be linked together with high relationship weight, high retention status, low adjusted acceptance resistance status, and current re-visit timing suitability.

[0249] The above customized prediction module (500) can determine not only customer-unit predictions but also the operational status of medical staff. For example, if medical staff B performed 3 lifting procedures and 2 laser procedures consecutively during the morning hours on the same day, accumulated a high total number of equipment shots, repeated movement between the examination room and the procedure room, and repeated the same explanation to multiple customers, the above customized prediction module (500) can adjust the basic exhaustion status upward based on the equipment load status, the number of space switchings, and the accumulated time of repeated explanations.

[0250] According to the above process 4, the customized prediction module (500) may additionally reflect claim-related voice peaks and a number of cases exceeding a set limit. For example, if a high-pitched claim-related utterance is detected during a consultation with a specific customer around 15:00, and immediately afterward the procedure is delayed by 18 minutes beyond the scheduled end time and the next slot is compressed, the customized prediction module (500) may switch the basic exhaustion state to an early warning stage or a higher warning stage even if it is in the stage immediately preceding the warning.

[0251] Subsequently, the visualization linkage module (600) can receive the exhaustion status and display the current status of medical staff B on the operation dashboard screen. For example, the visualization linkage module (600) can reduce the display of available slots for new reservations after 16:00 or output a status change request linked to the reservation management system. In addition, the visualization linkage module (600) can also provide operation notifications such as “recommendation for medical staff B to rest” or “need to review additional assignment.”

[0252] Conversely, if a continuous rest block without patient tags is secured from 16:20 to 16:45 and no additional calls or equipment interventions occur during that time, the custom prediction module (500) can alleviate the burnout state by reflecting the uninterrupted rest block. As a result, the visualization linkage module (600) can partially release the previously reduced reservation slots or display the additional assignment availability status for subsequent time slots again.

[0253] In the integrated embodiment described above, the data collection module (100) collects structured data and unstructured data regarding consultation cases, and the unstructured data normalization module (200) can convert the data into a structured code system. The medical management ontology construction module (300) connects the structured data to entities and relationships, and can adjust the strength of the relationship between specific skin concerns and procedures through Process 1. The data lifecycle flow management module (400) determines whether the data has been maintained sequentially from the collection stage to the prediction and visualization stage, and can assign a preservation status through Process 2. The customized prediction module (500) performs customer-specific predictions based on the relationship strength and the preservation status, and can determine the resistance to acceptance of the next-ranked customized proposal through Process 3. Additionally, the customized prediction module (500) can determine the medical staff load status within the same date according to the process 4, and the visualization linkage module (600) can display the prediction result and the operation status in conjunction with the dashboard screen, the reservation management screen, and the warning screen.

[0254] Accordingly, the judgment regarding consultation and procedure proposals for a specific customer and the judgment regarding schedule adjustments at the medical staff level do not operate as separate, fragmented functions, but can be processed in a continuous, connected manner within the same system. For example, if the relationship weight is high, the flow preservation state is good, and the acceptance resistance state is corrected to be low for the aforementioned customer A, the next-ranked customized proposal can be displayed as a priority review target. On the other hand, if the burnout state of medical staff B remains at a higher warning level at the same time, the visualization linkage module (600) can display the proposal itself but also present an operational judgment to adjust the actual timing of subsequent consultation assignment to after the next available slot. In this way, the present invention can be utilized to support management decision-making in medical institutions by mutually linking the customer-level recommendation flow and the medical staff-level operational flow.

[0255] Although preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above. Various modifications are possible by those skilled in the art without departing from the essence of the invention as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present invention. Explanation of the symbols

[0256] Data collection module (100) Unstructured data normalization module (200) Medical management ontology construction module (300) Data lifecycle flow management module (400) Customized prediction module (500) Visualization linkage module (600)

Claims

Claim 1 In a customized management decision support system for medical institutions, a data collection module (100) for collecting structured data and unstructured data inside and outside the medical institution; an unstructured data normalization module (200) for converting unstructured records collected or converted by the data collection module (100) into a structured data structure; a medical management ontology construction module (300) for creating a data structure in the form of a knowledge graph by defining entities related to medical institution management and relationships between entities based on the data standardized by the unstructured data normalization module (200); a data lifecycle flow management module (400) for tracking the processing flow of the structured data, the unstructured data, the structured data, and the knowledge graph data from the collection stage to the stages of conversion, normalization, relationship linkage, prediction reflection, and visualization reflection; and using relationship information generated by the medical management ontology construction module (300) and time-series flow data generated by the data lifecycle flow management module (400) for customer segmentation, revisit pattern analysis, assessment of customer churn potential, procedure demand analysis, or A customized prediction module (500) that generates next-ranked customized proposal data;and includes a visualization linkage module (600) that provides the results calculated by the customized prediction module (500) in the form of an operation dashboard screen, a report screen, a notification screen, or an operation review screen; the data collection module (100) collects unstructured data including consultation voice data and handwritten pen chart image data, and generates text data by performing voice recognition processing on the consultation voice data and optical character recognition processing on the handwritten pen chart image data; the unstructured data normalization module (200) extracts expressions related to the procedure name, procedure site, dosage, purpose of consultation, price, or sentiment from the text data and maps them to a standard code system; the medical management ontology construction module (300) generates relationship information by connecting entities related to customers, inflow channels, skin concerns, procedures, payments, and revisits; the data lifecycle flow management module (400) determines whether the relationship information is maintained until a subsequent stage along with the original source identifier; and the customized prediction module (500) provides a predicted value or based on the relationship information and the time series flow data. A medical institution-customized management decision support system that calculates a recommended value, and the visualization linkage module (600) outputs the predicted value or recommended value together with operational judgment reference information. Claim 2 In claim 1, the data collection module (100) is linked with at least one of a customer relationship management system, an electronic medical record system, a reservation management system, a sales management system, a wearable badge unit, and a handwritten pen chart input means to collect at least one of reservation data, sales data, consultation voice data, and pen chart image data; obtains a reservation date and time, reservation channel, reservation procedure name, reservation manager, and reservation change history from the reservation data; obtains a payment date and time, payment amount, sales item by procedure, refund status, and discount information from the sales data; links a consultation start time, consultation end time, consultation manager identification information, customer identification information, and consultation session identification information to the consultation voice data; links a shooting time, author identification information, customer identification information, and consultation session identification information to the pen chart image data; assigns at least one of source identification information, collection time information, customer identification information, consultation identification information, and medical treatment identification information to each of the collected data to enable linkage of identical data between subsequent modules; performs voice recognition processing on the consultation voice data to generate a voice transcription; and the pen chart image data A management decision support system tailored for medical institutions, which performs optical character recognition processing to generate a character extraction result, stores the original file and the voice transcription or character extraction result in a form that allows mutual mapping, checks for the omission of essential items, file corruption, or duplicate collection, and stores the check results as status values ​​indicating normal, suspected omission, candidate for duplicate, waiting for conversion, or need for re-collection, and provides them as criteria for subsequent processing. Claim 3 In claim 2, the unstructured data normalization module (200) performs a preprocessing step of performing at least one of spacing correction, special character removal, abbreviation expansion, synonym integration, error correction, and unit expression unification on voice transcription, optical character recognition result, free input memo, and counselor input text received from the data collection module (100); an item extraction step of extracting a procedure name, procedure site, dosage, purpose of consultation, expression of dissatisfaction, expression of inducing a revisit, price-related expression, and sentiment-related expression from the preprocessed text; a code mapping step of mapping the extracted items to a standard procedure code, site code, dosage code, sentiment classification code, or price inquiry code; and, in cases where multiple procedure names and multiple sites are written together in a single sentence or multiple text fragments within the same consultation session, separates the correspondence relationship of procedure-site-dosage using a syntactic analysis rule or a learning model; merges voice transcriptions, pen chart records, and consultation memos belonging to the same consultation session to reconstruct them into a single integrated structured record; and a procedure name abbreviation, equipment name, used within the medical institution, A management decision support system tailored for medical institutions that normalizes multiple expressions with the same meaning into the same standard item by referring to an internal dictionary including part expressions and capacity expressions, thereby enabling the use of information scattered across different media as consistent input values ​​in subsequent ontology construction and prediction stages. Claim 4 In claim 3, the medical management ontology construction module (300) defines customer, counselor, medical staff, visit route, skin concern, procedure name, procedure site, dosage, price, consultation sentiment, reservation conversion status, and revisit cycle as entities or attributes, and forms a knowledge graph path by connecting records generated in the reservation stage, consultation stage, treatment stage, payment stage, and revisit stage in chronological order based on the customer identification information, consultation identification information, procedure identification information, and occurrence time information, and after first assigning a basic weight to the relationship between a specific skin concern entity and a specific procedure entity, the number of overlapping occurrences representing the number of matching events in which the same procedure is identified together in voice transcription and pen chart records within the same consultation session, the emphasis marking density calculated by dividing the number of underline, asterisk, circle, or box mark objects formed in the area where the specific procedure is written in the pen chart by the normalized area of ​​the said area, and the echo utterance representing the number of times the same procedure name or the same efficacy expression appears repeatedly in the customer utterance within a preset time interval immediately after the medical staff's explanation. A medical institution-tailored management decision support system that collects the number of times, a synchronization delay time representing the difference between the time when a specific procedure is first mentioned in a verbal explanation by medical staff and the time when the procedure is first displayed on a chart, and the number of times modified entries including overwriting, duplicate marking, and deletion and re-entry for the same chart area, up-corrects the basic weight based on the number of overlapping appearances, the emphasis marking density, and the number of echo utterances, down-corrects the basic weight based on the synchronization delay time and the number of modified entries, stores the correction results as dynamic relationship weights which are attribute values ​​of knowledge graph edges connecting skin concern entities and procedure entities, and provides the dynamic relationship weights as input values ​​to distinguish between priority review relationships and subordinate relationships in a subsequent prediction stage. Claim 5 In claim 4, the medical management ontology construction module (300), in correcting the dynamic relationship weight, determines that if the synchronization delay time is within a preset first standard value, the voice description and chart display are linked within the same consultation flow and increases the basic weight by a first increment range; if the number of overlapping occurrences is greater than or equal to a preset second standard value and simultaneously the emphasis marking density is greater than or equal to a preset third standard value, performs a synergy correction greater than when a single condition is satisfied; adds the weight in stages as the number of echo utterances increases; if the synchronization delay time exceeds the first standard value, expands the reduction range in stages according to the degree of excess; if the number of modification entries is greater than or equal to a preset fourth standard value, determines that there is repeated modification for the same chart area and further lowers the weight; if the chart display time exists before the voice mention time and the synchronization delay time is calculated as a negative number, determines this as a preset template event and excludes it from the synergy correction target; and the weight is preset A management decision support system tailored for medical institutions that performs range limit control to prevent exceeding an upper limit, thereby reducing situations where relationship strength is excessively biased by specific input values, and provides the resulting edge attribute values ​​as criteria for distinguishing between priority and subordinate relationships in a subsequent prediction stage. Claim 6 In claim 5, the data lifecycle flow management module (400) assigns an original source identifier to each data, comprising at least one of an original voice file identifier, an original image file identifier, a customer identifier, a consultation session identifier, and a conversion task identifier; calculates a source chain completeness ratio using the ratio of the number of processing steps actually connected without interruption among the collection step, text conversion step, normalization mapping step, ontology linkage step, prediction reflection step, and visualization reflection step to the total number of processing steps; calculates the number of nodes in which the original source identifier is maintained without interruption among data nodes that have reached the final prediction reflection step or visualization reflection step as the number of valid nodes; identifies records in which the mapping with customer identification information, consultation session identification information, or the original source identifier is broken as orphan records; calculates a timestamp discontinuous accumulation value by accumulating the excess of the interval exceeding a preset allowable threshold among the time difference between adjacent processing steps; calculates the number of times the identifier is recovered according to an automatic recovery rule when a partial omission of identifiers or a failure in connection between steps occurs as the number of automatic recovery successes; and the operator directly reassigns the identifier or mapping Calculate the number of times relationships have been modified as the number of manual intervention corrections, assign one of a plurality of preservation status grades to the data set or the data node, classify the preservation status into multiple stages such as Excellent, Good, Average, Caution, and Reprocessing Required, and if the number of valid nodes is greater than or equal to a preset valid node threshold and simultaneously the source chain integrity ratio is greater than or equal to a preset chain integrity threshold, classify the data flow preservation status upward by a greater extent than when a single condition is satisfied, and if orphan records occur or the cumulative value of timestamp discontinuities exceeds a preset discontinuity cumulative threshold, classify the flow preservation status downward, provided that if the number of automatic recovery successes is greater than the number of manual intervention corrections, the downward extent is limited.A medical institution-customized management decision support system that ensures that data with maintained original sources and data with severed or missing sources are treated differently in subsequent processing steps by ensuring that data classified as the best or excellent grade is prioritized for input into the customized prediction module (500), and that data classified as caution or reprocessing required grade is moved to a reprocessing queue or output with a caution mark through the visualization linkage module (600).

Citation Information

Patent Citations

  • A System of Providing Hospital Management Consulting Service

    KR101917852B1

  • System and method for constructing ontology information system of Oriental medicine data

    KR1020200022109A

  • Hospital consulting provision system, hospital consulting provision method, and program recorded on recording medium

    KR1020220043797A

  • Method for constructing an integrated medical information database using a large language model and method for utilizing thereof

    KR102855049B1

  • Method for providing optimized solutions of medical institution management based on artificial intelligence

    KR102931134B1