Apparatus and method for medical data correlation analysis
The medical data correlation analysis device addresses the challenge of integrating AI diagnostic results by normalizing and analyzing relationships between AI outputs, enhancing clinical decision-making with standardized reporting and improved accuracy.
Patent Information
- Application Number
- KR1020250185147
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-11-06
- Estimated Expiration
- 2045-11-28
AI Technical Summary
Current medical AI platforms fail to provide comprehensive clinical guidance by correlating and interpreting diagnostic results from multiple AI solutions, leading to limited support for integrated clinical decision-making, especially in complex cases.
A medical data correlation analysis device that collects, normalizes, and analyzes AI diagnostic results to generate integrated clinical guides by calculating relationship reliability scores and generating reports based on predefined threshold criteria, supporting clinical decision-making.
Enhances clinical decision-making by integrating AI results from different manufacturers, reducing interoperability issues, and providing clear clinical guidance through standardized reporting, improving accuracy and transparency.
Smart Images

Figure 112025134235706-PAT00004_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to medical data processing technology, and more specifically, to technology for analyzing relationships between different AI diagnostic results and generating clinical guides. Background Technology
[0002] Recently, medical artificial intelligence (AI) technology capable of detecting patient abnormalities or diagnosing diseases based on various medical imaging data, such as CT, MRI, and X-ray, is rapidly advancing in the medical field. AI solutions specialized for specific diseases—such as lung nodule detection, cerebral hemorrhage detection, and cardiovascular disease analysis—are being widely adopted in hospitals and medical checkup centers, playing a crucial role in improving the reading efficiency and diagnostic accuracy of medical staff.
[0003] In addition, medical AI platforms are emerging to provide multiple AI solutions developed by different manufacturers within a single environment. By presenting various AI results on a single screen, these platforms enable medical institutions to utilize multiple AI solutions in an integrated manner. Medical professionals can view diagnostic results from different AI solutions through a single platform, improving accessibility and convenience compared to traditional diagnoses based on individual solutions.
[0004] However, current medical AI platforms are largely limited to simply listing and displaying the diagnostic results of each AI in parallel. For example, even if a lung nodule detection AI suggests "nodule detection" and a cardiovascular analysis AI suggests "risk of stenosis" for the same patient, the platform does not provide any correlation, clinical relevance, or comprehensive risk interpretation between the two results. Since medical professionals must interpret each result individually, comprehensive diagnostic support based on information from multiple AIs is limited; furthermore, it is difficult to make sufficient clinical judgments based solely on individual AI results, particularly in situations involving multi-institutional diseases or complex clinical patterns.
[0005] Due to these limitations, medical AI platforms require technology capable of going beyond a simple listing of results to collect and normalize multiple AI outcomes, analyze the relationships between AIs, and provide clinically significant guidance or test recommendations based on those associations. In other words, rather than medical professionals independently evaluating the outputs of each AI solution one by one, there is a need for a function that systematically analyzes the mutual reinforcement, consensus, and trade-off relationships between AIs and interprets them in an integrated manner. The problem to be solved
[0006] According to one embodiment, a medical data correlation analysis device and method are proposed to more effectively support the clinical decision-making of medical staff by collecting multiple medical AI diagnostic results in a standardized manner, quantitatively analyzing the correlations between AIs, and generating an integrated report and diagnostic guide with clinical value based on the analysis results. means of solving the problem
[0007] A medical data correlation analysis device according to one embodiment includes an AI result collection unit that receives and stores a plurality of AI diagnostic results from an AI platform, a normalization unit that converts the data format and attributes of the AI diagnostic results into a standardized analysis format, and a correlation analysis unit that calculates basic indicators for the normalized plurality of AI diagnostic results, determines the type of relationship between AIs to generate relationship labels, calculates a relationship reliability score using the relationship labels and basic indicators, and generates a clinical guide according to a predefined threshold criterion based on the reliability score.
[0008] The above AI result collection unit may include an AI result receiving unit that receives a plurality of AI diagnostic results from an AI platform, an AI result storage unit that stores the received AI diagnostic results together with metadata, and an AI result classification unit that classifies the stored AI diagnostic results by patient unit, test unit, or AI type unit.
[0009] The above normalization unit may include a data conversion unit that converts AI diagnostic results of different structures into an internal standard format, and an attribute standardization unit that standardizes metadata and diagnostic attributes included in the AI diagnostic results according to a consistent standard.
[0010] The above association analysis unit may include a data mapping unit that sorts and maps multiple normalized AI diagnostic results into the same patient case based on patient ID, examination date and time, or image UID.
[0011] The above data mapping unit can perform mapping of lesion units based on the ROI coordinates of the image-based AI diagnosis result, the bounding box of the lesion area, or the segmentation mask.
[0012] The above association analysis unit may include a feature extraction unit that extracts the presence or absence of a lesion, probability values, risk scores, lesion size, length, or area, ROI coordinates, mask-based shape information, or inspection-based numerical values.
[0013] The above-mentioned correlation analysis unit may include a basic indicator calculation unit that calculates a basic indicator representing the relationship between multiple AI diagnosis results, a relationship type determination unit that determines the type of relationship between multiple AI diagnosis results based on the basic indicator, and a reliability calculation unit that calculates a relationship reliability score by reflecting data reliability information in the determined relationship type and the basic indicator.
[0014] The above reliability calculation unit can correct the reliability score by applying at least one of a sample size weight, a model uncertainty weight, an image quality or missing rate-based weight.
[0015] A medical data association analysis device characterized by the association analysis unit further including a clinical guide generation unit that generates a clinical guide corresponding to the threshold criterion when the relationship reliability score satisfies a predefined threshold criterion.
[0016] The above association analysis unit may include a template mapper that maps a determined clinical guide to either a patient template or a physician template.
[0017] The medical data correlation analysis device may further include a report generation unit that generates a report for a patient or doctor including the clinical guide, the result of the relationship analysis, and a summary of the individual AI diagnosis result.
[0018] The medical data association analysis device may further include a feedback learning unit that collects the final reading results or clinical findings entered by medical staff and reflects them in the relationship model or policy correction of the association analysis unit.
[0019] A medical data correlation analysis method performed by a computing device according to another embodiment comprises: a step of receiving and storing a plurality of AI diagnostic results from an AI platform; a normalization step of converting the data format and attributes of the received AI diagnostic results into a standardized analysis format; a step of calculating a basic indicator representing the relationship between the plurality of AI diagnostic results for the normalized plurality of AI diagnostic results; a step of determining the type of relationship between the plurality of AI diagnostic results based on the basic indicator and generating a relationship label; a step of calculating a relationship reliability score using the relationship label and the basic indicator; and a step of generating a clinical guide according to the threshold criterion when the relationship reliability score satisfies a predefined threshold criterion. Effects of the invention
[0020] According to the present invention, unlike the existing method of simply listing multiple AI diagnostic results in parallel on a medical AI platform, the following effects can be obtained by collecting and normalizing multiple AI diagnostic results in a standardized manner and analyzing their interrelationships.
[0021] The present invention provides an environment that can normalize and manage multiple AI diagnostic results provided by different manufacturers into an integrated format, thereby resolving interoperability issues caused by differences in data formats, transmission protocols, and metadata structures between AI products. This significantly reduces the burden of complex customization or separate interoperability development required for each manufacturer in existing platforms.
[0022] This invention does not merely list the results individually generated by each AI, but rather directly analyzes and quantifies clinical relationships, such as correlation, reinforcement, and conflict, between AIs. Through this, medical professionals can interpret diagnostic signals from multiple AIs in an integrated manner and identify multi-organ diseases, complex patterns, and combinations of increased risk that could not be detected by the results of a single AI alone.
[0023] This invention supports medical professionals in taking clear clinical actions based on multiple AI outputs by automatically generating clinical guidance, such as recommendations for precise examinations, suggestions for alternative tests, and the necessity of re-interpretation, based on the analysis of relationships between AIs. This enables actual clinical decision-making support functions that were not available on existing platforms.
[0024] The present invention provides a consistent analysis pipeline, including the calculation of basic indicators, the generation of relationship labels, and the calculation of confidence scores, for comparing and evaluating the results of different AI solutions, thereby establishing consistent interpretation criteria for inconsistencies or combined effects between different AI solutions. This enhances the ability to detect result conflicts or unexpected patterns between AIs.
[0025] This invention can automatically generate reports for patients and physicians, allowing medical staff to intuitively view structured AI result summaries and relationship analysis results, while patients receive AI results in an easy-to-understand format. This contributes to enhancing the transparency and reliability of the AI-based medical process.
[0026] The present invention progressively improves analysis accuracy during platform operation by continuously calibrating and training an association analysis model based on feedback from medical professionals. This enables the continuous assurance of data-driven stability and clinical reliability during long-term use. Brief explanation of the drawing
[0027] FIG. 1 is a drawing illustrating the overall configuration of a medical data correlation analysis device according to an embodiment of the present invention. FIG. 2 is a diagram illustrating the detailed configuration of the AI result collection unit of FIG. 1 according to an embodiment of the present invention. FIG. 3 is a drawing illustrating the detailed configuration of the normalization unit of FIG. 1 according to an embodiment of the present invention. FIG. 4 is a diagram illustrating the detailed configuration of the association analysis unit of FIG. 1 according to an embodiment of the present invention. FIG. 5 is a diagram illustrating the detailed configuration of the correlation calculation unit of FIG. 4 according to an embodiment of the present invention. FIG. 6 is a diagram illustrating the overall flow of the association calculation process according to one embodiment of the present invention. FIG. 7a is a diagram showing an example of applying the process of a basic indicator calculation unit according to an embodiment of the present invention to three scenarios for calculating basic statistical indicators for different pairs of AI diagnostic results. FIG. 7b is a diagram illustrating the process of a relationship type determination unit determining a relationship type for each scenario according to one embodiment. FIG. 7c is a diagram illustrating a series of processes in which a reliability calculation unit according to an embodiment of the present invention calculates raw scores of consensus, reinforcement, and conflict based on basic indicators and relationship labels, calculates a final reliability grade by applying a sigmoid transformation, and then triggers a clinical guide according to a policy pack. FIG. 8 is a drawing illustrating the detailed configuration of the report generation unit of FIG. 1 according to an embodiment of the present invention. FIG. 9 is a drawing illustrating the detailed configuration of the feedback learning unit of FIG. 1 according to an embodiment of the present invention. FIG. 10 is a diagram illustrating the overall flow of a medical data correlation analysis method according to one embodiment of the present invention. Specific details for implementing the invention
[0028] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but can be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Throughout the specification, the same reference numerals refer to the same components.
[0029] In describing the embodiments of the present invention, if it is determined that a detailed description of known functions or configurations may unnecessarily obscure the essence of the invention, such detailed description will be omitted. Furthermore, the terms described below are defined considering the functions in the embodiments of the present invention, and these may vary depending on the intentions or conventions of the user or operator. Therefore, such definitions should be based on the content throughout this specification.
[0030] Additionally, each block or each step may represent a module, segment, or part of code containing one or more executable instructions for executing specific logical functions, and it should be noted that in some alternative embodiments, the functions mentioned in the blocks or steps may occur out of order. For example, two blocks or steps described in succession may actually be performed substantially simultaneously, and the blocks or steps may also be performed in the reverse order of the corresponding functions as needed.
[0031] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. However, the embodiments of the present invention exemplified below may be modified in various different forms, and the scope of the present invention is not limited to the embodiments described below. The embodiments of the present invention are provided to more completely explain the present invention to those skilled in the art to which this invention pertains.
[0032] FIG. 1 is a diagram illustrating the overall configuration of a medical data correlation analysis device according to one embodiment of the present invention.
[0033] The medical data correlation analysis device (hereinafter referred to as the ‘device’) (1000) according to the present invention receives a plurality of AI diagnostic results from an external AI platform (1010), collects and normalizes these results, analyzes their correlations, and then performs the function of generating a guide report that can be used for clinical decision-making based on the analysis results.
[0034] The AI platform (1010) is an independent AI execution system that operates in a hospital or cloud environment and corresponds to an 'AI diagnosis result providing unit' that receives image and examination data, generates only diagnosis results (probability value, risk score, lesion location, etc.) for each AI model, and provides them to the device (1000).
[0035] The medical data source (1020) is a configuration that provides various types of original medical data generated within the hospital. This includes imaging data such as CT / MRI / X-ray, clinical and health checkup data, and pathology, blood, and biochemical test data. The medical data source (1020) is an independent input source that provides the original medical data itself, rather than the 'result value' generated by the AI platform (1010).
[0036] Specifically, the medical data source (1020) includes detailed data groups such as an image data source (1021), a medical examination / EMR-based data source (1022), and a pathology / test data source (1023).
[0037] An image data source (1021) provides image data obtained from CT (Computed Tomography), MRI (Magnetic Resonance Imaging), X-ray, and other medical imaging equipment to an AI platform (1010). This image data is used as a basic input for performing image-based diagnosis, such as lesion detection, location analysis, and shape analysis, in each AI model.
[0038] The examination and EMR-based data source (1022) includes a health examination server, a hospital EMR (Electronic Medical Record) system, PACS, biosignal monitoring equipment, etc., and provides non-imaging diagnostic information such as a patient's clinical history, test results, and risk indicators to the AI platform (1010). This data forms a complementary relationship with the image-based AI model and can be used as input for multiple AI models by the platform.
[0039] The pathology and test data source (1023) provides pathological and biochemical values generated from blood test analysis equipment, pathology slide analysis equipment, automated testing devices, etc., to the AI platform (1010). The data includes quantitative information important for medical decision-making, such as the activity of lesions, inflammatory response, and metabolic status.
[0040] Various medical data output from the medical data source (1020) configured in this manner is transmitted to the AI platform (1010), and the AI platform (1010) executes multiple AI models (AI-1, AI-2, AI-3, etc.) based on this to generate respective diagnostic results. The generated AI diagnostic results are subsequently collected and normalized by the device (1000) and used as input for correlation analysis between multiple AI diagnostic results.
[0041] The AI platform (1010) receives patient data from a medical data source (1020) and generates each AI diagnosis result through multiple AI models (AI-1, AI-2, AI-3 …). The AI platform (1010) can integrate and provide multiple AI analysis solutions with different manufacturers, models, and algorithms, and each AI diagnosis result is transmitted to the device (1000) in different data formats such as DICOM SR, HL7, JSON, and XML.
[0042] The device (1000) may be implemented by one or more computing devices, and its components may operate according to program instructions executed by a processor. That is, the device (1000) may be implemented in the form of a server, a workstation, or a cloud-based system, and the function of each component described below may be implemented as a hardware circuit or realized as a software or firmware module executed by a processor.
[0043] The device (1000) of the present invention may include an AI result collection unit (110), a normalization unit (120), an association analysis unit (130), a report generation unit (140), and a feedback learning unit (150).
[0044] The AI result collection unit (110) receives various AI diagnostic results generated from the AI platform (1010) and introduces them into the system. The AI result collection unit (110) includes an interface function that can accept data of different formats transmitted from the AI platform (1010) in a consistent input form, and the collected results are stored and classified for subsequent analysis. The detailed components and operation process of the AI result collection unit (110) are described in more detail through FIG. 2, which will be described later.
[0045] The normalization unit (120) converts the data format and attributes of the AI diagnostic results provided by the AI platform (1010) into a standardized analysis format. Since the data structure or diagnostic value representation method output by each AI is different, it converts them into a unified common analysis schema to make them suitable as input for subsequent correlation analysis. The detailed components and operation process of the normalization unit (120) are explained in more detail through FIG. 3.
[0046] The association analysis unit (130) calculates a basic indicator representing the relationship between multiple normalized AI diagnostic results, determines the type of relationship between AIs, calculates a relationship reliability score, and generates one or more clinical guides according to a predefined threshold criterion based on the reliability score.
[0047] When determining the type of relationship between AIs, the association analysis unit (130) determines mutual reinforcement relationships, consensus relationships, conflicting relationships, precedence / delay relationships, spatial coincidence relationships, etc. between AIs and extracts the clinical value of the result in a quantified form. The detailed components and operation process of the association analysis unit (130) are explained in more detail through FIGS. 4 and FIGS. 5.
[0048] The report generation unit (140) organizes information into a report format that can be provided to medical staff or patients based on the clinical guide and analysis results of the association analysis unit (130), and visualizes it. The report includes clinical guide information such as a summary of individual AI diagnosis results, results of relationship analysis between AIs, and test recommendations or follow-up plans based on the type of relationship. The detailed components and operation process of the report generation unit (140) are explained in more detail through FIG. 8.
[0049] The feedback learning unit (150) collects reading results or clinical findings input by medical staff and reflects them in the form of relearning in the analysis model and policy of the correlation analysis unit (130). By doing so, it performs the role of gradually improving the accuracy and clinical suitability of the device (1000). The detailed components and operation process of the feedback learning unit (150) are explained in more detail through FIG. 9.
[0050] FIG. 2 is a diagram illustrating the detailed configuration of the AI result collection unit of FIG. 1 according to an embodiment of the present invention.
[0051] The AI result collection unit (110) performs preprocessing to stably introduce multiple AI diagnostic results generated from the AI platform (1010) into the device. The AI result collection unit (110) may include an AI result receiving unit (111), an AI result storage unit (112), and an AI result classification unit (113).
[0052] The AI result receiving unit (111) receives AI diagnostic results generated by various AI solutions linked to the AI platform (1010) in real time. Since the data format and transmission protocol of the AI diagnostic results differ depending on the manufacturer, algorithm, and analysis field, the AI result receiving unit (111) is designed to support various communication methods and data formats.
[0053] Communication methods include, for example, DICOM network protocol, REST API calls, SDK-based integration, and message queue-based asynchronous transmission.
[0054] Data formats include, for example, DICOM SR structured reports, HL7 messages, JSON response data, XML-based structured data, etc.
[0055] The AI result receiving unit (111) includes an interface layer for integrally processing these multiple protocols and is configured so that structural information such as a patient identifier, image UID, and analysis time can be extracted together when the result is received. Through this, heterogeneous AI diagnostic results are stably accepted as input to the device's internal analysis pipeline.
[0056] The AI result storage unit (112) stores the AI diagnosis results received through the AI result receiving unit (111) in a database (DB). When storing, not only are the diagnosis values stored, but essential metadata is also recorded so that each result can be accurately sorted and mapped in a subsequent step. This metadata may include patient identification information (Patient ID), examination date and time (Study / Series time), image UID (e.g., DICOM SOP Instance UID), AI model type (AI-1, AI-2, etc.), and diagnosis value (probability, score, positive / negative label, etc.). This serves as a basis for matching AI diagnosis results to the same patient and same examination unit in a subsequent correlation analysis step.
[0057] In addition, the AI result storage unit (112) can perform functions such as queuing and batch storage so that continuous and unstructured data can be processed without loss even if it is repeatedly input.
[0058] The AI result classification unit (113) performs the function of organizing the stored AI diagnosis results into an analyzable structure. Since the AI models operate with different disease domains and purposes, the data must be grouped according to consistent criteria during the analysis phase.
[0059] The AI result classification unit (113) can perform automatic classification by patient unit, examination unit, or AI type. Patient unit classification is classifying all AI diagnostic results for the same patient into one group. Examination unit classification is classifying results generated at the same time of the same image or examination. AI type classification is classifying according to the diagnostic field of the model, such as lung CT-based AI, brain MRI-based AI, and cardiovascular risk AI.
[0060] Through this function, diagnostic results generated by multiple AIs for one patient can be systematically arranged by test, model, and time, and a basis is established for quantitatively calculating the relationship between them in the correlation analysis unit (130).
[0061] In addition, the AI result classification unit (113) can store model-specific metadata (institution, manufacturer, version information, etc.) together and be used for subsequent learning or quality correction that takes into account domain differences.
[0062] FIG. 3 is a diagram illustrating the detailed configuration of the normalization unit of FIG. 1 according to an embodiment of the present invention.
[0063] Referring to FIG. 3, the normalization unit (120) includes a data conversion unit (121) and an attribute standardization unit (122) to refine various AI diagnostic results input from the AI platform (1010) into an internal standard analysis format.
[0064] The data conversion unit (121) converts the diagnostic results transmitted from the AI platform (1010) into a standard analysis format within the device. The AI diagnostic result formats provided for each AI solution are highly heterogeneous, such as DICOM SR, HL7 messages, JSON, and XML, and because the attribute structure, field names, and nesting methods differ for each format, they cannot be used as input for association analysis as they are. The data conversion unit (121) interprets these unstructured structures and reformats them into a structure that fits the standard DB model defined internally. For example, the diagnostic result of a lung nodule AI provided in JSON has fields such as model name, probability value, and lesion coordinates, and the data conversion unit (121) maps these to a standard field system (e.g., lesion_presence, lesion_probability, roi_center). On the other hand, since the diagnostic result of a DICOM SR-based cerebral hemorrhage AI has a tree structure, the data conversion unit (121) parses the lesion location, lesion size, and diagnostic phrase from the SR node and maps them to the same standard analysis format. This conversion process may include parsing syntax of different formats (JSON / XML / SR / HL7, etc.), replacing manufacturer-specific representations with internal standard keys, removing fields not needed for analysis, and error-tolerant correction processes in the event of missing fields or format mismatches.
[0065] As such, the data conversion unit (121) is responsible for converting the "appearance (structure)" of the data into a completely unified standard analysis format, which serves as the first foundation for enabling the data to be handled in the subsequent attribute standardization stage and the association analysis unit (130).
[0066] The attribute standardization unit (122) standardizes the metadata and diagnostic attributes included in the AI diagnostic results converted into a unified standard analysis format by the data conversion unit (121) according to a consistent standard. Because AI models express patient information, test reference values, diagnostic codes, and result values in different ways, the notation method differs from model to model even for attributes that have the same meaning. The attribute standardization unit (122) eliminates these differences so that all AI diagnostic results can be compared on an equal baseline.
[0067] Specifically, the attribute standardization unit (122) can unify patient identification information. It unifies Patient ID formats that differ from hospital system to hospital system, and also converts time information such as Study Date / Acquisition Time using the same rules. In addition, it accurately maps image reference IDs (e.g., Study Instance UID, Series UID, SOP Instance UID) so that AI diagnostic results for the same test are reliably matched.
[0068] The attribute standardization unit (122) can standardize clinical codes and lesion names into standard terms. It converts the diagnostic codes or lesion names output by each AI into a standard terminology system such as ICD-10, or unifies lesion names with the same meaning. For example, "lung_nodule", "L-nodule", "suspected nodule" → is standardized to "Lung Nodule".
[0069] The attribute standardization unit (122) can normalize the scale of the result values (probability, score, grade). It normalizes values that differ from the output rules of each AI model to a common scale. For example, a score of 0 to 100 is converted to a scale of 0 to 1, High / Medium / Low is normalized to a 3-level risk value, and Positive / Negative is unified to a binary value of 1 / 0. This normalization of result values is essential because the same standard values are required when calculating the correlation coefficient, odds ratio, MI, ΔP, etc. in the association analysis unit (130).
[0070] FIG. 4 is a diagram illustrating the detailed configuration of the association analysis unit of FIG. 1 according to an embodiment of the present invention.
[0071] The association analysis unit (130) receives multiple normalized AI diagnosis results, analyzes the relationship between each AI diagnosis result within the same patient unit according to statistical and clinical criteria, and converts the analysis results into guide information having clinical significance.
[0072] The association analysis unit (130) may include a data mapping unit (131), a feature extraction unit (132), an association calculation unit (133), and a clinical guide generation unit (134), and each component operates sequentially and hierarchically to enable the interpretation of various clinical patterns such as correlations, reinforcements, trade-offs, and precedences between AI diagnosis results.
[0073] The data mapping unit (131) arranges multiple tests performed on the same patient and results generated by multiple AI models into a single case. Since the normalized AI diagnostic results are generated at different times, different modalities, and different models, the multiple results are rearranged into a single patient-based (one-case, multi-AI set) based on basic reference information such as patient ID, test date and time, and image UID.
[0074] In addition, the data mapping unit (131) enables mapping of lesion units by utilizing location information such as the ROI (Region of Interest) center coordinates, the bounding box of the lesion area, and the segmentation mask in the case of image-based AI diagnosis results. Through this, a structure is provided that allows meaningful clinical events of the same patient to be grouped and analyzed even when the modalities are different, such as lung nodule AI and cardiovascular risk AI.
[0075] The feature extraction unit (132) extracts key features required for association analysis from each mapped AI diagnosis result. The extracted features may include quantitative features such as the presence or absence of a lesion (positive / negative), a lesion probability value or risk score, lesion size, length, area, etc., the location of the lesion (ROI center coordinates, mask-based location information), shape information (segmentation-based shape descriptor), and test-based numerical values (z-score, intensity, count, etc.).
[0076] The feature extraction unit (132) extracts meaningful clinical units (features) from data provided by different AI models in a normalized form, thereby enabling statistical comparison in the association calculation step described later.
[0077] The association calculation unit (133) quantifies the relationship between multiple AI diagnostic results based on features extracted within the same patient case and labels the relationship type. The association calculation unit (133) can calculate the association based on indicators such as Pearson / Spearman correlation coefficient, MCC (Matthews Correlation Coefficient), odds ratio (log-OR and 95% CI), mutual information (MI), conditional probability difference (ΔP), AUROC or risk-based indicators, and ROI-based agreement (IoU, Dice, central distance). These indicators are calculated as normalized values so that they can be compared even when different types of information (imaging, examination, pathology) are combined, and consequently, an association index (range 0-1) is calculated.
[0078] The association calculation unit (133) can generate not only simple correlations but also labels for semantically classifying clinical relationships between AI diagnosis results. These labels may include relationship types such as Agreement, Complementary, Contradictory, Leading, ROI-based spatial match, conditional pattern, rare combination / abnormal pattern.
[0079] Agreement refers to multiple AIs presenting the same findings. Complementary refers to heterogeneous information, such as imaging AIs and diagnostic AIs, reinforcing each other. Contradictory refers to a specific AI presenting a positive result while another presents a negative result. Leading refers to one AI's risk signal appearing temporally earlier. ROI-based spatial agreement means two AIs identify the same lesion region. Conditional patterns have a significant association only under specific conditions, such as in the elderly or those with underlying diseases. Rare combinations and abnormal patterns are meaningful patterns with low incidence rates.
[0080] The clinical guide generation unit (134) generates a clinical guide corresponding to the threshold when the relationship reliability score satisfies a predefined threshold. For example, if the results of the lung nodule AI and the cardiovascular risk AI diagnosis results rise together, a "recommendation for cardiovascular precision examination" can be generated as a clinical guide; if the positive findings of the cerebral hemorrhage AI and the hypertension risk pattern are combined, a "blood pressure management guideline" can be generated; if the two AI diagnosis results conflict, a "re-examination and alternative examination recommendation" can be generated; and if a lesion in the same area is confirmed based on ROI, a "follow-up examination for the corresponding lesion" can be generated as a clinical guide.
[0081] The clinical guide generation unit (134) determines the results based on pre-set policy rules for each relationship type and can convert them into phrases or test recommendation items suitable for patient / doctor reports.
[0082] FIG. 5 is a diagram illustrating the detailed configuration of the correlation calculation unit of FIG. 4 according to an embodiment of the present invention.
[0083] The association calculation unit (133) multidimensionally analyzes the relationship between multiple AI diagnosis results within the same patient case, calculates a relationship type label and a final reliability score based on objective indicators, and selects and outputs a policy applicable to clinical decision-making.
[0084] The association calculation unit (133) may include a basic indicator calculation unit (1331), a relationship type determination unit (1332), a reliability calculation unit (1333), and a template mapper (1334). These components operate sequentially to extract various relationship types, such as agreement, reinforcement, conflict, precedence, and spatial match among multiple AIs, and calculate the clinical reliability of each relationship as a continuous value in the range of 0 to 1.
[0085] The base metric calculation unit (1331) calculates base metrics to quantitatively express the relationship between multiple AI diagnosis results, and the base metrics are used as key input indicators for the correlation calculation step and as input values for the relationship type determination unit (1332) and the reliability calculation unit (1333).
[0086] As key basic indicators, the Pearson / Spearman correlation coefficient signifies a monotonic, linear relationship between continuous values, such as lesion probabilities and risk scores. The Matthews Correlation Coefficient (MCC) represents the degree of agreement between positive / negative based AIs. The odds ratio (log-OR) and 95% confidence interval (CI) represent the combined risk between binary outputs. Mutual Information (MI) signifies the amount of information shared between heterogeneous modalities (imaging and examinations). The Conditional Probability Difference (ΔP) represents the increase or decrease in B positivity when A is positive. Other key basic indicators include partial correlation, covariate correction indicators, and ROI-based spatial indicators (IoU, DICE, centroid distance).
[0087] The calculated basic indicators are organized into an internally structured dictionary (e.g., r, mcc, log_or, mi, delta_p, auroc, cramers_v, etc.) and can subsequently be used for determining relationship labels and calculating confidence.
[0088] The relationship type determination unit (1332) receives basic indicators and determines the relationship type between the AI diagnosis results. The relationship type is designed to be clinically meaningful beyond simple correlation.
[0089] As examples of representative relationship types, Agreement (or Cross-validation) means that two AIs present the same conclusion (e.g., both positive); Complementary means that image-based AI and test-based AI reinforce each other; and Contradictory means clinical discrepancy, such as AI A being positive and AI B being negative. Leading means that the risk signal from AI A occurs temporally before that of AI B; Sensitivity-Specificity Complement refers to the structure of screening AI and confirmatory AI; and Synergy means that accuracy increases when using AI A and AI B in combination compared to using them alone. Conditional (Subgroup) means a strong relationship only in specific groups such as the elderly group or the underlying disease group, ROI Match means that two AIs identify the same lesion location, Lagging means that AI B is confirmed Δt hours after AI A, and Rare combo means a rare AI combination pattern with p < 1%.
[0090] Each relationship type can be determined based on critical conditions. For example, if MCC ≥ 0.30, it is determined as 'Consensus'; if r ≥ 0.60 or MI ≥ 0.20, as 'Reinforcement'; if OR ≤ 0.67 or ΔP ≤ -0.10, as 'Conflict'; if CCF peak lag ≤ 72h & Granger p<0.05, as 'Leading'; and if IoU ≥ 0.30 or Dice ≥ 0.50, as 'Spatial Match'.
[0091] The output of the relationship type determination unit (1332) is a relationship label (relation_type ∈ {agreement, complementary, contradictory, ...}), and the relationship becomes the input signal of the confidence calculation unit (1333).
[0092] The reliability calculation unit (1333) calculates a reliability score (score, 0 to 1) that can be directly used for clinical decision-making by integrating basic indicators and quality and domain information for relationship labels (e.g., consensus, reinforcement, conflict, precedence, etc.) generated by the relationship type determination unit (1332), and performs the role of activating a specific guide or policy trigger according to a predefined threshold when necessary.
[0093] The reliability calculation unit (1333) first collects the elements constituting the input values. The input may include relationship labels and basic indicators (Pearson / Spearman coefficients, MI, MCC, OR / ΔP, AUROC, IoU / Dice, sample size N, p-value, CI, etc.). Furthermore, w_model based on model uncertainty and entropy, w_quality reflecting image quality, noise, missing rate, etc., and w_domain reflecting deviation between equipment and sites (ECE / Brier / random effect, etc.) may also be included. In addition, policy and environment information such as version control (e.g., threshold_pack_id, calib_pack_id, site_id, device_id) is also collected, and if missing values exist, correction-based processing (using winsorize or masked) is used instead of simple removal.
[0094] Subsequently, sample size correction coefficient during the preprocessing stage (Example N0=50) is applied, total weight It is calculated. In cases of significant model uncertainty or domain differences between hospitals, the corresponding weight is lowered, having a conservative effect on the final score; conversely, in reliable data environments, the raw score is reflected more directly.
[0095] Next, different raw score calculation formulas are applied depending on the relationship type. For agreement, the raw score is calculated by combining MCC, AUROC deviation(max(AUROC-0.5, 0)), correlation coefficient r, etc., and an example formula can be defined as raw_agree = 1.5·MCC + 0.3·r + 2·max(AUROC-0.5, 0). For complementary, it has a structure such as raw_boost = 1.3·r + 1.0·ρ + 1.2·min(MI,1), and if there is a space-based relationship, IoU or Dice values are additionally included. Contradictory consists of a formula reflecting negative correlation and risk reduction, and can reflect raw_conflict = -2·min(r,0) + 2·max(-MCC,0) + (1-min(OR,1)) + 1.5·max(-ΔP,0), etc. In the case of lead / lag, the raw_lead relationship, which combines the cross-correlation (CCF) peak and the Granger causality score, is used, for example, in the form raw_lead = z(CCF_peak) + β·1(p_granger < 0.05) - γ·(|Δt-Δt0| / Δt0).
[0096] The raw scores calculated in this way are normalized into final confidence levels ranging from 0 to 1 through a sigmoid transformation, and the scores are generally
[0097]
[0098] It is implemented in the form of. The calculated reliability scores are classified according to grading criteria as "Very High" for 0.90 or higher, "High" for 0.75 or higher, "Medium" for 0.60 or higher, and "Low" for values lower than that, providing reliability grades that clinicians can intuitively judge.
[0099] The reliability calculation unit (1333) can activate policy triggers according to predefined threshold packs. For example, if the agreement reliability score is 0.70 or higher, a detailed examination may be recommended; if the boost reliability score is 0.75 or higher, an additional examination or rapid follow-up examination may be recommended; and if the conflict is 0.70 or higher, a re-reading or alternative examination workflow may be automatically triggered. Additionally, the policy structure may have a priority order of Conflict > Agreement > Boost > Other, and the output result includes template_id (for patient / doctor), order_set (examination / follow-up order), etc., which are transmitted to the report generation unit (140).
[0100] Additionally, the reliability calculation unit (1333) can correct the reliability score by applying at least one of a sample size weight, a model uncertainty weight, an image quality or missing rate-based weight.
[0101] The template mapper (1334) combines the relationship labels and reliability scores (0–1) generated by the relationship type determination unit (1332) and the reliability calculation unit (1333), as well as patient, equipment, and time context information, to select and determine the final clinical guide according to a predefined policy pack and safety priority rules.
[0102] The template mapper (1334) combines and produces one or more clinical guides with clinical significance, such as recommendations for detailed examination, additional examination, re-reading or alternative examination, follow-up schedules, display abbreviation / hiding, and placement of attention, according to the type of relationship (e.g., consensus, reinforcement, conflict, precedence, etc.). Additionally, the determined clinical guides are output by rendering different text templates and order sets according to the role (for patients / for doctors), and all decision processes are recorded as metadata along with version information and managed to enable subsequent analysis and reproduction.
[0103] The template mapper (1334) receives relationship labels and reliability scores from the relationship type determination unit (1332) and the reliability calculation unit (1333), and prioritizes reviewing cases where a safety gate is required, such as conflicting relationships or rare patterns, and applies a threshold-based defense policy. Subsequently, one or more clinical guides are selected according to priority rules regarding consensus, reinforcement, spatial match, precedence, etc., and the selected clinical guides are mapped to patient or physician templates and corresponding test order sets to produce a final output. The determined policy and template information is stored along with version information and managed to enable subsequent auditing and reproducibility.
[0104] FIG. 6 is a diagram illustrating the overall flow of the association calculation process according to one embodiment of the present invention.
[0105] More specifically, Figure 6 shows a step-by-step analysis procedure that receives diagnostic results independently generated by multiple AI models as input, calculates basic statistical indicators, determines relationship labels, calculates confidence scores, determines threshold-based policy triggers, and generates a final report.
[0106] First, the device (1000) receives multiple AI diagnostic results from the AI platform (1010) (S601). In this step (S601), along with basic information for mapping patient units, test units, and lesion units, results structured and standardized through the normalization unit (120) are provided as input to the association analysis unit (130).
[0107] Next, the device (1000) calculates base metrics (S602). In this step (S602), various predefined statistical and spatial metrics, such as Pearson-Spearman correlation coefficient, MCC, odds ratio (log-OR), mutual information (MI), ΔP, AUROC-based metrics, and ROI-based spatial similarity (IoU, Dice), are calculated to derive base metrics that serve as the basis for relationship analysis.
[0108] When the basic indicator is calculated, the device (1000) receives the basic indicator and determines the relationship labels between multiple AIs (S603). For example, if two AI models present the same opinion, a relationship label such as 'agreement' is determined; if image-based and non-image-based AIs reinforce each other, a relationship label such as 'complementary' is determined; if the conclusions between AIs conflict, a relationship label such as 'contradictory' is determined; and if a temporal precedence relationship exists, a relationship label such as 'lead' is determined. The determination of the relationship is based on whether specific indicators such as MCC, r, MI, OR, and ΔP satisfy predefined threshold conditions.
[0109] Next, the device (1000) receives relationship labels and basic indicators and calculates a relationship confidence score (S604). In this step (S604), the device (1000) calculates a total weight w_{total} by reflecting quality factors such as model uncertainty, image / data quality, domain bias (ECE / Brier), and sample size weight, and then generates an initial score by applying a raw score formula for each relationship type. Subsequently, the value is normalized into a continuous value between 0 and 1 through a sigmoid function, and as a result, each relationship type is expressed as a score-based confidence level (e.g., very high / high / medium / low).
[0110] The device (1000) receives a normalized reliability score and activates a final guide trigger according to threshold conditions of a predefined policy pack (S605). For example, if the consensus reliability score is 0.70 or higher, a recommendation for a detailed inspection may be triggered; if the reinforcement reliability score is 0.75 or higher, an additional inspection or short-term tracking may be triggered; and if the conflict reliability score is 0.70 or higher, a re-reading or alternative inspection may be triggered, respectively.
[0111] Next, the device (1000) performs a policy-template mapping step (S606) based on the activated trigger. In this step (S606), the device (1000) determines which policy is selected (e.g., recommendation for detailed examination, conflict flag, follow-up examination), what form of report text the policy will be converted into, which of the patient template and the doctor template will be applied, and what the examination order set mapped to the selected policy is.
[0112] The policy pack defines report text templates (e.g., plain-language description for patients, clinical guide for physicians), test recommendation items, risk level expressions, warning display rules, etc., corresponding to each policy type (agree / boost / conflict), and the device (1000) generates structured report components by selecting and replacing appropriate templates according to the activated trigger and clinical scenario.
[0113] After the policy and template mapping is completed, the device (1000) generates a final report based on this (S607). The report includes a summary of individual AI diagnosis results, relationship analysis results, a reliability score, and selected guidelines (recommendation for detailed examination, follow-up schedule, conflict warnings, etc.), which can be directly used for clinical decision-making by medical staff.
[0114] FIG. 7a is a diagram showing an example of applying the process of a basic indicator calculation unit according to one embodiment of the present invention to three scenarios for calculating basic statistical indicators for pairs of different AI diagnostic results.
[0115] The base metric calculation unit (1331) applies an appropriate statistical technique based on the combination of continuous and binary values between the two input AI diagnostic results to generate base metrics for determining the relationship type in the relationship type determination unit (1332).
[0116] Figure 7a includes the results of calculating actual baseline indicators for the following three clinical cases.
[0117] (Scenario 1) Pneumonia pA ↔ CRP
[0118] Analysis of the association between pneumonia probability AI (pA) and CRP positivity revealed a low positive correlation with a Pearson correlation coefficient of 0.141 and a Spearman correlation coefficient of 0.133. The MCC was 0.043, indicating a weak agreement close to zero, while the OR was 1.202, which is slightly greater than 1 but insufficient to signify a statistically strong effect. The ΔP was +0.015, showing only a weak change where CRP positivity increased by 1.5%p when pA was positive.
[0119] These basic indicators indicate a weak association with only a weak positive (+) trend overall.
[0120] (Scenario 2) ICH AI ↔ Coagulation abnormality
[0121] The baseline indicators comparing the AI diagnosis results of ICH (cerebral hemorrhage) detection with the blood coagulation abnormality results are MCC -0.369, OR 0.214, and ΔP -0.230, all of which deviate significantly in the negative direction. This signifies a strong negative (-) discrepancy signal, indicating that the two results lead to opposite conclusions.
[0122] (Scenario 3) CACS ↔ CAD Risk
[0123] In the association analysis between CACS and CAD risk AI, very high levels of positive correlation are observed, with Pearson at 0.834, Spearman at 0.788, and MI at 0.596. This indicates high shared information, where the two values increase in strong synchronization.
[0124] FIG. 7b is a diagram illustrating the process of a relationship type determination unit determining a relationship type for each scenario according to an embodiment of the present invention.
[0125] The relationship type determination unit (1332) generates a relationship label for each relationship type based on a threshold condition.
[0126] (Scenario 1) Pneumonia pA ↔ CRP - Supportive Evidence
[0127] The underlying indicators of the pA-CRP pair do not meet the criteria for determining reinforcement, such as r ≥ 0.6, ρ ≥ 0.6, and MI ≥ 0.2, nor do they satisfy the consensus criteria such as MCC and OR. Therefore, this relationship does not fall under any label (consensus, reinforcement, or conflict) and is classified as a 'supporting basis' level with a weak positive tendency.
[0128] (Scenario 2) ICH AI ↔ Coagulation Abnormality - Contradictory Relationship
[0129] This relationship satisfies all conflicting criteria, such as OR ≤ 0.67, ΔP ≤ -0.1, and MCC ≤ -0.2. Therefore, the relationship type determination unit (1332) determines that the two AI diagnostic results are in a conflicting relationship in which they present opposite conclusions.
[0130] (Scenario 3) CACS ↔ CAD Risk - Complementary Relationship
[0131] In this case, all reinforcement criteria, such as r ≥ 0.6, ρ ≥ 0.6, and MI ≥ 0.2, are satisfied. Therefore, this relationship is classified as a complementary relationship that reinforces each other's information.
[0132] FIG. 7c is a diagram illustrating a series of processes in which a reliability calculation unit according to an embodiment of the present invention calculates raw scores of consensus, reinforcement, and conflict based on basic indicators and relationship labels, calculates a final reliability grade by applying a sigmoid transformation, and then triggers a clinical guide according to a policy pack.
[0133] The reliability calculation results are calculated as follows for each of scenarios 1, 2, and 3.
[0134] (Scenario 1) Pneumonia pA ↔ CRP
[0135] With a weight of w=0.769 applied, the consensus raw=0.392→score=0.575 (low), reinforcement raw=0.316→score=0.561 (low), and conflict raw=0→score=0.500 (low), with all three confidence levels falling below the threshold. Therefore, this relationship is classified as having a weak association and is classified as a supporting evidence level, and it does not activate policy triggers.
[0136] (Scenario 2) ICH AI ↔ Coagulation abnormality
[0137] A weight of w=0.767 is applied, and a strong conflict confidence is calculated with conflict raw=1.869→score=0.807 (high). This is a case where OR, ΔP, and MCC all satisfy the conflict condition, and recommendations for re-interpretation or alternative image inspection are activated according to the policy pack.
[0138] (Scenario 3) CACS ↔ CAD Risk
[0139] A weight of w=0.812 is applied, and reinforcement reliability is highest with raw=2.586 → score=0.891 (high). This relationship can trigger high-risk guidance, such as recommendations for cardiac close examinations, depending on the policy pack (boost≥0.75).
[0140] FIG. 8 is a diagram illustrating the detailed configuration of the report generation unit of FIG. 1 according to an embodiment of the present invention.
[0141] Referring to FIG. 8, the report generation unit (140) is configured to generate reports for patients and doctors based on the results of association analysis and reliability calculation, and may include a result summary unit (141), an association analysis visualization unit (142), and a guide recommendation unit (143).
[0142] The result summary section (141) organizes and displays the diagnostic results independently produced by each AI model in a consistent format on the report. This is a function that organizes and provides to the user results, such as prediction probabilities, risk scores, and abnormalities, provided by models of different domains, such as image-based AI, test-based AI, and biosignal-based AI, in a standardized format. For example, if the chest X-ray AI produces "suspected pneumonia (0.82)" and the blood test AI produces "increased inflammation level (0.79)," the result summary section (141) lists the two results together so that the user can intuitively grasp the overall diagnostic signal.
[0143] The association analysis visualization unit (142) expresses the relationships between AIs in various visual forms, such as graphs, network diagrams, bar charts, and relationship matrices, based on the relationship labels (cross-validation, reinforcement, conflict) and confidence scores calculated by the association analysis unit (130). For example, if a relationship is determined to be complementary, the connection line between the two AIs is thick and displayed in a color indicating a positive (+) direction. If a relationship is determined to be conflicting, a warning indicator is visualized along with a color indicating a negative (-) direction. This visualization helps patients and medical staff intuitively understand the structural relationships between different AI diagnosis results.
[0144] The guide recommendation unit (143) recommends clinical actions suitable for a specific patient (recommendation for detailed examination, re-examination, additional examination, alternative imaging examination, etc.) based on the consensus / reinforcement / conflict score calculated by the reliability calculation unit (1333) and the threshold criteria defined in the policy pack.
[0145] Policy packs may include the following criteria.
[0146] Conflict ≥ 0.70 → Re-interpretation and alternative imaging examination recommended
[0147] Boost ≥ 0.75 → Recommendation for evidence-based additional testing for high-risk reinforcement
[0148] agree ≥ 0.70 → Suggest strengthening of grounds for suspecting the same disease, priority given to detailed examinations
[0149] The recommended guide is reflected in the patient report and the physician report, respectively.
[0150] In the case of patient reports, examples include "AI1 (Chest X-ray): Suspected pneumonia, AI2 (Blood Test): Increased inflammation levels → The two results are consistent, making pneumonia highly probable," and "AI diagnostic results are inconsistent, requiring further verification." Patient reports are structured to exclude technical jargon and instead guide patients through the relationships between results (consistency, inconsistency, or reinforcement) and clinically necessary actions using concise language.
[0151] For physician reports, examples include "AI1 confidence 0.82, AI2 confidence 0.79 → interpreted as a confounding relationship (additional CT / echo recommended)." and "A lesion was observed on AI3 (CT), but the corresponding abnormality was not confirmed on AI4 (MRI), so it was determined to be a conflicting relationship → re-imaging or alternative test required within 24 hours." Physician reports display confidence, relationship labels, baseline indicators such as ΔP, OR, and MCC, and policy pack-based recommendation guides, and include detailed information to enable clinicians to make actual clinical decisions.
[0152] FIG. 9 is a diagram illustrating the detailed configuration of the feedback learning unit of FIG. 1 according to an embodiment of the present invention.
[0153] Referring to FIG. 9, the feedback learning unit (150) may include a medical staff feedback collection unit (151) and a model learning unit (152).
[0154] The medical staff feedback collection unit (151) provides an API (Application Programming Interface) that can be linked with a UI (User Interface) or a hospital information system (HIS / RIS) to support medical staff in inputting or automatically linking medical staff feedback information. The medical staff feedback information may include ground truth (whether the actual confirmed diagnosis is true), clinical findings (whether the lesion is confirmed, location of the lesion, extent, etc.), whether it matches or does not match the AI diagnosis result, clinical judgment in case of conflict (re-examination, alternative test, etc.), and results of follow-up tests after treatment (optional).
[0155] The collected feedback information is used as ground truth data to verify the clinical validity of the label, relationship type (consensus, reinforcement, conflict), and relationship reliability for each AI diagnostic result, according to actual clinical standards.
[0156] The model learning unit (152) performs retraining to continuously improve the relationship model of the association analysis unit (130) based on the collected medical staff feedback data.
[0157] The model learning unit (152) structures and stores medical staff feedback together with the existing analysis dataset and checks consistency between relationship labels and basic indicators. The calculation rules (weights, thresholds, etc.) for conflict, reinforcement, and consensus reliability scores can be corrected based on clinical data.
[0158] The model learning unit (152) reflects new feedback information as retraining data according to a set period (e.g., daily, weekly, monthly) and causes the statistical model or ML-based relationship model of the association analysis unit (130) to gradually improve.
[0159] FIG. 10 is a diagram illustrating the overall flow of a medical data correlation analysis method according to one embodiment of the present invention.
[0160] Each module of the device (1000) is interconnected to provide a complete structure for a series of processes including the collection, normalization, relationship analysis, labeling, reliability calculation, clinical guide generation, report output, and retraining based on medical staff feedback.
[0161] Input step (S1001)
[0162] The device (1000) receives patient data from a medical data source (1020), such as imaging equipment (CT, MRI, X-ray), pathology and blood test equipment, and a health checkup server, and receives AI diagnostic results for the patient data from an AI platform (1010) or an AI solution within the hospital. The input data may include not only original medical data such as image information and numeric test values, but also structured AI diagnostic results expressed in JSON, DICOM SR, etc.
[0163] AI diagnosis result collection stage (S1002)
[0164] The device (1000) receives diagnostic results generated by an external AI engine or an AI solution within the hospital through an AI result collection unit (110) in the form of DICOM, REST API, SDK, message queue, etc., and classifies and stores them by patient unit and test unit.
[0165] Normalization step (S1003)
[0166] The device (1000) converts different AI models and manufacturer-specific output formats into an internal standard schema through the normalization unit (120), and standardizes key attributes such as patient ID, image UID, and examination date. Probability values, risk scores, etc. are also unified into a common unit.
[0167] Association analysis step (S1004)
[0168] The device (1000) calculates basic indicators (Pearson, Spearman, MCC, OR, MI, ΔP, etc.) through the association analysis unit (130).
[0169] Labeling and reliability calculation step (S1005)
[0170] The relationship label is determined according to the above basic indicator and predefined classification rules, and the device (1000) calculates a relationship raw score by reflecting data reliability information in the relationship label and basic indicator, and calculates a reliability score by applying a sigmoid transformation.
[0171] Reliability scores for each relationship label are calculated by reflecting the stability of the relationship, data quality, and agreement between models. For example, the reliability of a reinforcement relationship increases when there is a high MI and correlation coefficient, and the reliability of a trade-off relationship increases when OR·ΔP·MCC strongly detaches in the negative direction.
[0172] Guide generation step (S1006)
[0173] The reliability calculation results are entered into a policy pack, and threshold-based clinical guidelines (recommendations for precise testing, need for re-interpretation, follow-up testing, etc.) are generated.
[0174] Conflict ≥ 0.70 → Re-interpretation or alternative imaging examination recommended
[0175] boost ≥ 0.75 → Reinforcement evidence-based high-risk guide
[0176] agree ≥ 0.70 → Consensus-based fine scan recommended
[0177] Report generation step (S1007)
[0178] The device (1000) generates a patient / doctor report by synthesizing correlation information between the determined guide and the AI through the report generation unit (140).
[0179] Feedback learning step (S1008)
[0180] The device (1000) can register actual final diagnosis results and clinical judgment information based on reports from medical staff as feedback, and utilizes the feedback information as retraining data through the feedback learning unit (150).
[0181] The present invention has been described above with reference to its embodiments. Those skilled in the art will understand that the present invention may be implemented in modified forms without departing from the essential characteristics of the invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the invention is defined by the claims, not by the foregoing description, and all variations within the scope of the claims should be interpreted as being included in the invention.
Claims
Claim 1 A medical data association analysis device comprising: an AI result collection unit that receives and stores multiple AI diagnosis results from an AI platform; a normalization unit that normalizes the data format and attributes of the multiple AI diagnosis results into a standardized analysis format; and an association analysis unit; wherein the association analysis unit comprises: a data mapping unit that aligns and maps the normalized multiple AI diagnosis results to the same patient case and performs mapping at the lesion unit level; a basic indicator calculation unit that calculates a basic indicator representing the relationship between multiple AI diagnosis results; a relationship type determination unit that determines at least one relationship type among agreement, reinforcement, conflict, precedence, or spatial coincidence relationships between the multiple AI diagnosis results based on the basic indicator and generates a relationship label; a reliability calculation unit that calculates a relationship reliability score by reflecting data reliability information in the determined relationship type and the basic indicator; and a clinical guide generation unit that generates a clinical guide corresponding to the threshold when the relationship reliability score satisfies a predefined threshold. Claim 2 A medical data correlation analysis device according to claim 1, wherein the AI result collection unit comprises: an AI result receiving unit that receives a plurality of AI diagnostic results from an AI platform; an AI result storage unit that stores the received AI diagnostic results together with metadata; and an AI result classification unit that classifies the stored AI diagnostic results by patient unit, test unit, or AI type unit. Claim 3 A medical data correlation analysis device according to claim 1, wherein the normalization unit comprises: a data conversion unit that converts AI diagnostic results of different structures into an internal standard format; and an attribute standardization unit that standardizes metadata and diagnostic attributes included in the AI diagnostic results according to a consistent standard. Claim 4 A medical data correlation analysis device according to claim 1, wherein the data mapping unit sorts and maps a plurality of normalized AI diagnostic results into the same patient case based on patient ID, examination date and time, or image UID. Claim 5 A medical data association analysis device according to claim 4, wherein the data mapping unit performs mapping of lesion units based on ROI coordinates of an image-based AI diagnosis result, a bounding box of a lesion area, or a segmentation mask. Claim 6 A medical data association analysis device according to claim 1, wherein the association analysis unit comprises a feature extraction unit that extracts the presence or absence of a lesion, a probability value, a risk score, the size, length, or area of the lesion, ROI coordinates, mask-based shape information, or a test-based numerical value. Claim 7 delete Claim 8 A medical data association analysis device according to claim 1, wherein the reliability calculation unit corrects the reliability score by applying at least one of a sample size weight, a model uncertainty weight, an image quality or missing rate-based weight. Claim 9 delete Claim 10 A medical data association analysis device according to claim 1, wherein the association analysis unit comprises a template mapper that maps a determined clinical guide to one of a patient template or a physician template. Claim 11 The medical data association analysis device according to claim 1 further comprises a report generation unit that generates a report for a patient or doctor including the clinical guide, the relationship analysis result, and the summary of the individual AI diagnosis result. Claim 12 The medical data association analysis device according to claim 1 further comprises a feedback learning unit that collects final reading results or clinical findings entered by medical personnel and reflects them in the relationship model or policy correction of the association analysis unit. Claim 13 A method for analyzing medical data associations performed by a computing device, comprising: receiving and storing a plurality of AI diagnostic results from an AI platform; a normalization step of normalizing the data format and attributes of the plurality of AI diagnostic results into a standardized analysis format; aligning and mapping the normalized plurality of AI diagnostic results into the same patient case, while performing mapping at the lesion unit level; calculating a basic indicator representing the relationship between the plurality of AI diagnostic results; determining at least one type of relationship among agreement, reinforcement, conflict, precedence, or spatial coincidence relationships between the plurality of AI diagnostic results based on the basic indicator and generating a relationship label; calculating a relationship reliability score by reflecting data reliability information in the determined relationship type and the basic indicator; and, if the relationship reliability score satisfies a predefined threshold criterion, generating a clinical guide according to the threshold criterion.
Citation Information
Patent Citations
A clinical decision support ensemble system and the clinical decision support method by using the same
KR1020170140757A
Method for prcessing and acquiring pathology slide image and apparatur for performing the same
KR1020230150134A
Uncertainty Estimation in Medical Imaging
US20230368913A1
