Method and apparatus for supporting subject-customized prescription based on systemic disease risk

Abstract

A method and an apparatus for supporting a subject-customized prescription based on systemic disease risk are disclosed. An apparatus for providing subject-customized assistance information, according to one embodiment, comprises at least one processor, wherein the at least one processor: inputs medical data of a subject into a systemic disease-related risk prediction model so as to process the medical data; acquires, from the systemic disease-related risk prediction model, risk information about the subject as a processing result of the medical data; and provides the subject-customized assistance information on the basis of the risk information, the customized assistance information being generated using a risk change pattern analysis method for analyzing a change in the systemic disease risk over time on the basis of the risk information and / or a risk factor weight calculation method for calculating a weight of the risk information as one of a plurality of risk factors for cardiovascular disease.

Description

Method and device for supporting personalized prescriptions based on systemic disease risk

[0001] The present invention relates to the field of artificial intelligence (AI)-based medical image analysis technology, and more particularly, to a method and device for providing subject-tailored auxiliary information by utilizing systemic disease risk and / or cardiovascular risk factor information calculated through fundus image analysis.

[0002] With the recent advancement of medical image analysis utilizing artificial intelligence (AI), systems that predict and diagnose diseases based on medical images are being developed. In particular, AI models that predict the risk of systemic diseases such as cardiovascular disease (CVD) and chronic kidney disease (CKD) through fundus image analysis are beginning to be utilized in clinical settings.

[0003] Diabetes is becoming an increasingly common chronic disease, and with it, the importance of managing its various complications, including diabetic retinopathy, cardiovascular disease, and kidney disease, is growing. Traditionally, diabetic patients must visit ophthalmologists for the diagnosis and follow-up of diabetic retinopathy, and internal medicine specialists separately for the management of diabetes, cardiovascular disease, and kidney disease. This often results in inconvenience and hinders integrated management.

[0004] AI technology based on fundus images analyzes microvascular changes in the retina, demonstrating the potential to predict the risk of not only ophthalmic diseases but also systemic diseases such as diabetes, hypertension, cardiovascular disease, cerebrovascular disease, kidney disease, brain disease, neurological disease, liver disease, and pulmonary disease. Utilizing this technology, for example, internal medicine physicians could utilize fundus image analysis results to consider the likelihood of developing diabetic retinopathy and the risk of systemic diseases, helping them develop personalized diagnosis and treatment plans. This could reduce the burden of multiple department visits for patients and enable more integrated patient management for medical professionals.

[0005] However, conventional AI-based diagnostic assistance technologies have primarily focused on predicting disease risk or specific risk indicators and providing them to medical professionals as reference information. This approach suffers from a lack of clear guidelines or linkage systems for how to specifically link and utilize AI-provided predictive information in actual clinical practice to patient medication prescribing decisions. For example, when AI predicts a patient's high risk of cardiovascular disease, there is insufficient information on how to weight this prediction relative to existing risk factors (age, blood pressure, cholesterol levels, smoking status, diabetes, etc.) or how to interpret changes in risk over time to inform adjustments in medication (e.g., statins, antihypertensives, etc.) prescription intensity.

[0006] As a result, medical professionals only use AI analysis results as one of several references for clinical judgment, making it difficult to actively and objectively incorporate them into personalized medication prescribing decisions. Furthermore, there were limitations in effectively utilizing AI predictive information for assessing disease progression and treatment response, key aspects of chronic disease management. These issues hinder the full realization of the medical value of AI technology in clinical practice.

[0007] The technical problems of the present invention are as follows.

[0008] First, the systemic disease risk information generated by artificial intelligence can provide objective and useful auxiliary information that can be specifically utilized in actual drug prescription decisions.

[0009] Second, by analyzing the patient's long-term risk change trends, it can provide information that helps in early detection of accelerated disease progression or the possibility of complications and in determining the appropriate timing of therapeutic intervention.

[0010] Third, the risk information generated by artificial intelligence can be compared with existing traditional cardiovascular risk factors to assess their relative importance (weight), which can then provide a rational basis for adjusting the intensity of drug prescriptions.

[0011] Fourth, for patients with chronic diseases such as diabetes with multiple complications, more comprehensive and efficient patient management can be supported by integrating ophthalmological information and systemic disease risk information.

[0012] The technical problems of the present application are not limited to the problems described above, and problems not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present invention pertains from this specification and the attached drawings.

[0013]

[0014] A device for providing customized auxiliary information of a subject according to one embodiment includes at least one processor, wherein the at least one processor inputs medical data of the subject into a systemic disease-related risk prediction model to process the medical data, obtains risk information of the subject as a result of processing the medical data from the systemic disease-related risk prediction model, and provides customized auxiliary information for the subject based on the risk information, wherein the customized auxiliary information can be generated using at least one of a risk change pattern analysis method for analyzing a change in systemic disease risk over time based on the risk information, or a risk factor weight calculation method for calculating a weight of the risk information as one risk factor among a plurality of risk factors for cardiovascular disease.

[0015] The technical solutions of the present application are not limited to the above-described solutions, and solutions not mentioned can be clearly understood by a person skilled in the art to which the present invention pertains from this specification and the attached drawings.

[0016] The present invention can provide the following effects.

[0017] First, the present invention goes beyond simply providing information on the risk of systemic diseases predicted by artificial intelligence, and can support medical professionals to utilize it as an objective and specific basis for judgment when deciding on drug prescription by providing risk acceleration indicators or risk factor weightings with existing risk factors.

[0018] Second, monitoring the patient's long-term risk change patterns can contribute to improving prognosis by providing supplementary information to enable early detection of the possibility of complications or accelerated disease progression and to consider preemptive drug treatment intervention or prescription adjustment before clinical symptoms become apparent.

[0019] Third, by providing quantitative weighting information on the extent to which risk information generated by AI has influence compared to existing well-known risk factors (e.g., diabetes, smoking, hypertension, etc.), it can help medical professionals intuitively understand the clinical significance of AI prediction results and rationally reflect them in prescription decisions.

[0020] Fourth, especially in patients with diabetes who are at risk for multiple complications, a single fundus image can comprehensively assess not only ophthalmic status but also the risk of major systemic diseases such as cardiovascular and renal diseases and their changing trends, and consider these in prescriptions, thereby improving the efficiency and quality of patient management.

[0021] Fifth, from the patient's perspective, it becomes possible to comprehensively assess and manage their health status without having to visit multiple departments, and from the medical staff's perspective, it increases treatment efficiency by providing prescription assistance information based on objective indicators.

[0022] The effects of the present application are not limited to the effects described above, and effects not mentioned can be clearly understood by a person having ordinary skill in the art to which the present invention pertains from this specification and the attached drawings.

[0023]

[0024] Figure 1 is a drawing for explaining the configuration of a diagnostic device according to one embodiment.

[0025] Figure 2 is a diagram for explaining a risk prediction model according to one embodiment.

[0026] FIG. 3 is a diagram for explaining a method for analyzing a risk change pattern according to one embodiment and providing target-tailored auxiliary information using the same.

[0027] Figure 4 is an exemplary graph for explaining the meaning of a risk acceleration indicator according to one embodiment.

[0028] Figure 5 is a diagram for explaining the distribution of risk scores according to the duration of diabetes and the presence or absence of complications according to one embodiment.

[0029] FIG. 6 is a diagram for explaining a method for calculating risk factor weights according to one embodiment and a method for providing target-tailored auxiliary information using the same.

[0030] FIG. 7 is a diagram illustrating the usefulness of a cardiovascular risk score as a new biomarker according to one embodiment.

[0031] Figure 8 is a diagram for explaining an attention weight distribution according to one embodiment.

[0032] FIG. 9 is a drawing for explaining a method for providing target-tailored auxiliary information according to one embodiment.

[0033] A device for providing customized auxiliary information of a subject according to one embodiment includes at least one processor, wherein the at least one processor inputs medical data of the subject into a systemic disease-related risk prediction model to process the medical data, obtains risk information of the subject as a result of processing the medical data from the systemic disease-related risk prediction model, and provides customized auxiliary information for the subject based on the risk information, wherein the customized auxiliary information can be generated using at least one of a risk change pattern analysis method for analyzing a change in systemic disease risk over time based on the risk information, or a risk factor weight calculation method for calculating a weight of the risk information as one risk factor among a plurality of risk factors for cardiovascular disease.

[0034] The advantages and features of the present invention, and the methods for achieving them, will become clearer with reference to the embodiments described in detail below together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. These embodiments are provided solely 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 present invention, and the present invention is defined solely by the scope of the claims.

[0035] The terminology used herein is for the purpose of describing embodiments only and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise. As used herein, the terms "comprises" and / or "comprising" do not exclude the presence or addition of one or more other components in addition to the mentioned components. Like reference numerals refer to like components throughout the specification, and "and / or" includes each and any combination of one or more of the mentioned components. Although "first", "second", etc. are used to describe various components, these components are not limited by these terms. These terms are only used to distinguish one component from another. Therefore, it should be understood that a first component mentioned below may also be a second component within the technical spirit of the present invention.

[0036] Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used in their common sense to those skilled in the art to which the present invention pertains. Furthermore, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0037] In this specification, “image” may mean multi-dimensional data composed of discrete image elements (e.g., pixels in a two-dimensional image and voxels in a three-dimensional image).

[0038]

[0039] Basic configuration_Learning device and diagnostic device

[0040] Below, a diagnostic system that diagnoses a subject using medical data is described. In this specification, the term "diagnosis" may refer to a diagnostic aid for predicting health risks, rather than directly diagnosing a disease. Furthermore, in this specification, the term "diagnosis" may be used interchangeably with the term "prediction" in that it predicts health risks. For example, in this specification, a diagnostic system may be referred to as a prediction system, and a diagnostic device may be referred to as a prediction device. For convenience of explanation, the term "diagnosis" is used below, but it may also refer to a diagnostic aid and / or prediction.

[0041] In addition, the subject is the subject of health risk prediction information, and in this specification, the subject may be used with the same meaning as terms such as patient, predicted subject, and predicted subject patient.

[0042] Additionally, health risks may include at least one of cerebrovascular disease, kidney disease, neurological disease, liver disease, diabetes, hypertension, hypotension, and hyperlipidemia. Furthermore, health risks may include various diseases that may occur in the subject.

[0043] A diagnostic system may include a learning device and a diagnostic device. The learning device may perform model training. The model may be at least one machine learning and / or deep learning model. For example, the learning device may learn the model by performing preprocessing, training, validation, and testing using medical data. Furthermore, the diagnostic device may use the model trained by the learning device to predict the health risks of a subject.

[0044]

[0045] Figure 1 is a drawing for explaining the configuration of a diagnostic device according to one embodiment.

[0046] Referring to FIG. 1, the diagnostic device (10) may include a processor (12), a storage module (11), and a communication module (13).

[0047] The processor (12) may include one or more of a CPU (Central Processing Unit), a RAM (Random Access Memory), a GPU (Graphics Processing Unit), one or more microprocessors, and other electronic components capable of processing input data according to predetermined logic. In addition, there may be at least one processor (12).

[0048] The processor (12) can read the system program and various processing programs stored in the storage module (11). For example, the processor (12) can deploy processes, methods, etc. for performing the diagnosis described below on RAM and perform various processing according to the deployed programs. For example, the processor (12) can process the algorithm described in this specification. The processor (12) can predict health risks using a machine learning model. In addition, the processor (12) can implement various methods described in this specification.

[0049] The storage module (11) can store models. For example, the storage module (11) can store machine learning models, parameters and variables of the machine learning model, algorithms described herein, etc. In addition, the storage module (11) can store models learned by the learning device.

[0050] The storage module (11) can be implemented as a nonvolatile semiconductor memory, hard disk, flash memory, RAM, ROM (Read Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), or other types of tangible nonvolatile recording media.

[0051] The storage module (11) can store various processing programs, parameters for performing program processing, or data resulting from such processing. For example, the storage module (11) can store programs for performing the diagnosis described below, parameters, and data obtained by performing such programs. In addition, the storage module (11) can store various machine learning models described below.

[0052] Although not shown, the diagnostic device (10) may further include an input module. The input module may obtain user input. For example, the input module may obtain user input. In addition, the input module may obtain physical information of the subject described in this specification (at least one of height, weight, age, gender, race, smoking status, blood pressure (e.g., blood pressure level, high blood pressure), diabetes (or blood sugar level), and cholesterol level). In addition, the input module may obtain medical data (e.g., various medical images, various medical information, fundus images, MRI, etc.).

[0053] The communication module (13) can communicate with various devices (e.g., a learning device). For example, the diagnostic device (10) can receive a model learned in the learning device from the learning device through the communication module (13). For example, the communication module (13) can include a wired or wireless module that transmits data through a mobile communication module and / or other various communication standards.

[0054] Additionally, the diagnostic device (10) may be implemented in the form of a server. Additionally, the diagnostic device (10) may be implemented in the form of a client device.

[0055] In addition, the learning device may also include a storage module, a processor, and a communication module. At least some of the descriptions of the storage module (11), processor (12), and communication module (13) described above may be applied to the storage module, processor, and communication module of the learning device. For convenience of explanation, the descriptions of the storage module, processor, and communication module of the learning device are omitted.

[0056]

[0057] Basic configuration: data, model structure, and behavior

[0058] Medical data can include various forms of medical data. For example, medical data may include ophthalmology-related medical data, cardiovascular-related medical data, kidney (kidney), liver, brain, and lung and respiratory-related medical data.

[0059] For example, ophthalmology-related medical data may include retinal images, ocular images, fundus images, and optical coherence tomography (OCT). Cardiovascular-related medical data may include cardiac CT, cardiac MRI, cardiac ultrasound, carotid ultrasound, and intravascular ultrasound (IVUS). In addition, lung and respiratory-related medical data may include chest X-rays and chest CTs. Other medical data may include brain CTs and angiography.

[0060] Medical data can be labeled with multiple health indicators, which can be used to assess a subject's current condition related to a specific disease or to predict the risk of developing a specific disease in the future. Each of these multiple health indicators has clinical significance and serves as a crucial criterion for analyzing a subject's overall health from various perspectives.

[0061] Multiple health indicators may include cardiovascular health indicators, which can be divided into direct and indirect indicators. Direct indicators may include imaging biomarkers that reflect arteriosclerosis and general vessel status. These indicators include coronary artery calcification score (CAC score), carotid intima-media thickness (CIMT), and arterial plaque information. These indicators can be subcategorized as cardiovascular imaging biomarkers.

[0062] The coronary artery calcification score (CAS) measures the degree of calcification within the coronary arteries, directly assessing the degree of atherosclerosis. Carotid intima-media thickness (IMT) measures the thickness of the carotid artery wall, providing an indicator of arterial health. Furthermore, arterial plaque information analyzes detailed characteristics, such as plaque size, volume, composition, and stability, to assess vascular health. Additionally, retinal vascular characteristics can also be used as direct indicators. Characteristics observed in retinal images, such as tortuosity, diameter, and fractal dimension, provide important information related to cardiovascular health.

[0063] Indirect markers are non-imaging biomarkers used to complement the assessment of cardiovascular health. These include blood pressure, blood sugar levels, cholesterol levels, smoking status, HbA1c, age, sex, and body mass index (BMI). Blood pressure indirectly assesses cardiovascular health through systolic and diastolic blood pressure. Blood sugar levels, including fasting blood sugar and HbA1c, are used to assess diabetes and blood sugar control. Cholesterol levels, including total cholesterol, LDL, HDL, and triglyceride levels, play a key role in indirectly assessing cardiovascular risk. Smoking status is a risk factor for cardiovascular disease, and current and past smoking habits are used as indicators. Additionally, age, sex, body mass index (BMI), and family history of cardiovascular disease are also important indirect indicators of cardiovascular health.

[0064] Ocular diseases can also provide valuable information regarding cardiovascular disease. For example, diabetic retinopathy is an eye disease closely related to vascular health, providing indirect clues to systemic vascular health through the retinal vascular status in diabetic patients. Furthermore, retinal vein occlusion, retinal artery occlusion, hypertensive retinopathy, macular degeneration, retinal microangiopathy, optic nerve ischemic lesions, and retinal neovascularization provide information related to cardiovascular disease.

[0065] Additionally, multiple health indicators, along with medical data, are used as training data for risk prediction models (machine learning models), playing an important role in assessing the health status of subjects with specific diseases and predicting future risks.

[0066]

[0067] Figure 2 is a diagram for explaining a risk prediction model according to one embodiment.

[0068] Referring to FIG. 2, in (A), the processor of the diagnostic device can input medical data into a risk prediction model (100) and process the medical data using the risk prediction model (100) to generate risk information. In addition, the processor of the diagnostic device can generate subject-specific auxiliary information based on the risk information.

[0069] The risk prediction model (100) may include a first model (110) for predicting multiple health indicators for predicting an individual's health risk based on medical data, and a second model (120) for predicting an individual's health risk by analyzing the multiple health indicators generated thereby. The risk prediction model (100) may be included in a diagnostic device. For example, the risk prediction model (100) may be stored in a storage module of the diagnostic device, and a processor of the diagnostic device may load the risk prediction model (100) from the storage module and use the risk prediction model (100) to predict a subject's health risk.

[0070]

[0071] The model (110) receives medical data and generates multiple health indicators, and may be based on a neural network-based model. Specifically, the model (110) may be a neural network-based model that processes medical data (e.g., fundus images) to generate multiple health indicators. The processor may process the medical data using the model (100) to generate multiple health indicators predicted for the current health status of the subject whose medical data has been scanned. The multiple health indicators may include health indicators for multiple diseases that are highly correlated, such as cardiovascular health indicators (e.g., CAC score) and renal health indicators (e.g., eGFR). In addition, the processor may selectively extract health indicators included in a specific disease from among the multiple health indicators.

[0072] The Coronary Artery Calcium Score (CAC) is a quantitative measure of the degree of calcification in the coronary arteries and can be an important indicator for assessing the risk of cardiovascular disease. A CAC score of 0 indicates the absence of calcification, while a higher score may indicate an increased risk of coronary artery disease. For example, the CAC score is divided into 0 (no calcification), 1-100 (mild calcification), 101-300 (moderate calcification), 301-1000 (severe calcification), and >1000 (very severe calcification), and each range can have different clinical implications. Furthermore, the estimated Glomerular Filtration Rate (eGFR) is a key indicator of kidney function and can be an estimate of the amount of blood filtered by the glomeruli. The eGFR is usually expressed in units of ml / min / 1.73m², and a lower value may indicate decreased kidney function. eGFR can generally be classified into the following ranges: 90 or more (normal), 60-89 (mild decline), 30-59 (moderate decline), 15-29 (severe decline), and less than 15 (renal failure).

[0073] Cardiovascular disease, kidney disease, diabetes, and obesity are closely related conditions. These diseases share common pathological mechanisms—metabolic abnormalities, inflammation, and vascular and renal dysfunction—and the onset and progression of each disease impacts the others. To systematically understand this, the American Heart Association (AHA) proposed the Cardiovascular-Kidney-Metabolic (CKM) syndrome model. This model systematically describes disease progression stages and clarifies the interactions among cardiovascular disease, kidney disease, diabetes, and obesity. Specifically, it depicts a progression from the onset of metabolic risk factors and chronic kidney disease to the onset of potential cardiovascular disease, ultimately leading to the development of overt clinical cardiovascular disease. Simultaneously learning multiple health indicators related to these highly correlated diseases can effectively reflect the interconnectedness between these conditions.

[0074] More specifically, the model (110) may include a feature extraction module (111) and an output generation module (112).

[0075] First, the feature extraction module (111) can input medical data and extract key features of the input medical data. In one embodiment, the feature extraction module (111) can be configured as a common sub-model based on a neural network and can process the medical data to automatically learn clinically meaningful information. First, the processor can convert the medical data into a form suitable for processing by the feature extraction module (111) through preprocessing, such as pixel intensity adjustment, resolution change, and noise removal. Then, the feature extraction module (111) can utilize deep learning techniques such as convolutional neural networks (CNNs) to detect important patterns in the medical data, such as lesion patterns, vascular structures, and tissue densities, and convert these into data representations (feature vectors). The generated feature vectors are data representations that can be commonly used across multiple health indicators and can include important information required for predicting each indicator. This process contributes to increasing learning efficiency when learning or predicting multiple health indicators simultaneously, and can optimize model performance while saving computational resources.

[0076] Next, the output generation module (112) can receive the feature vector generated by the feature extraction module (111) as input and generate a prediction result for each health indicator. The output generation module (112) can include multiple head sub-models and can perform prediction optimized for the characteristics of each health indicator. Although the feature vector generated by the feature extraction module (111) includes common features, each health indicator has different clinical meanings and patterns, and therefore, a sub-model specialized for each indicator may be required to independently analyze and predict them.

[0077] The output generation module (112) can be configured so that multiple head sub-models can operate in parallel to simultaneously predict all health indicators. This maximizes learning and prediction efficiency and can derive optimal prediction results that reflect the specificity of each indicator while utilizing common feature representations. This allows for predictions of various health indicators within a single model structure, thereby ensuring both accuracy and clinical reliability. For example, the output generation module (112) can generate multiple predicted health indicators (e.g., cardiovascular health indicators, renal health indicators, blood pressure, etc.) related to diseases that are highly correlated with the subject's current health status. Furthermore, the output generation module (112) can selectively extract health indicators associated with specific diseases from among the multiple health indicators and output the extracted health indicators. For example, the health indicators to be extracted may be predetermined or may be input from an external source.

[0078] In addition, the extracted health indicators may include systemic disease risk information. The systemic disease risk information may include at least one of cardiovascular risk information, renal risk information, liver risk information, brain risk information, or pulmonary and respiratory risk information. For example, the systemic disease risk information may include cardiovascular risk information or renal risk information, or may include both cardiovascular risk information and renal risk information. In addition, the processor may apply weights to multiple detailed systemic disease risk information, such as cardiovascular risk information and renal risk information, to generate a single systemic disease risk information. For example, the processor may apply a first weight to the cardiovascular risk information, apply a second weight to the renal risk information, and generate a sum of the two as a single systemic disease risk information.

[0079]

[0080] In addition, in (B), the model (110) described in (A) may be composed of a first model (115) and a second model (120). The first model (115) may include the feature extraction module (111) and the output generation module (112) described above. Since the description of the first model (115) may be applied to the description of the model (110), the feature extraction module (111), and the output generation module (112) described in (A), a detailed description thereof will be omitted.

[0081] The second model (120) can be designed based on statistics, and can analyze the correlation between health indicators and health risks by utilizing various statistical techniques such as the Cox Proportional Hazards Model (hereinafter referred to as the Cox model) and logistic regression. The second model (120) can receive health indicators (e.g., cardiovascular health indicators, renal health indicators, blood pressure, etc.) output from the first model (115) as input values, analyze the relationship between them, and calculate health risks.

[0082] For example, the Cox model is a statistical technique suitable for survival analysis and can be used to assess the risk of developing major cardiovascular events (e.g., myocardial infarction, stroke) over a specific period. The Cox model mathematically analyzes the proportional risk relationship between health indicators and event occurrence, and can dynamically calculate the risk by including a time variable. That is, the output of the second model (120) is a result of a comprehensive assessment of an individual's health risk, and can be provided in the form of, for example, "risk of developing a major cardiovascular disease event within 5 years" or "probability of having a cardiovascular disease risk of 10% or higher."

[0083] Model (110) of (A), the first model (115) of (B), or the second model (120) can perform a more precise health risk analysis by combining not only health indicators derived from medical data, but also other personal data as additional inputs. Personal data that can be utilized as additional inputs includes various health-related information in addition to medical data. For example, personal data that can be utilized as additional inputs, i.e., additional data, may include clinical data such as age, gender, body mass index (BMI), smoking status, family history, blood pressure, and blood sugar levels. This additional data, along with health indicators derived from medical data, can contribute to a more comprehensive and accurate analysis of an individual's health risk.

[0084] Additionally, the first model (115) can utilize additional data as auxiliary input for health indicator prediction. For example, when a processor inputs blood pressure information along with a fundus photograph into the first model (115) to predict a CAC score, the blood pressure information can serve as an additional feature that considers its correlation with vascular status, thereby increasing the accuracy of the prediction.

[0085] The second model (120) can incorporate additional data into the health risk analysis process. For example, the processor can use health indicators such as the CAC score and eGFR generated from the first model (115) as main inputs for the Cox model of the second model (120), while additional data such as age, gender, and family history can be input into the second model (120) for integrated analysis. Through this combination, the processor can utilize the second model (120) to gain a deeper understanding of an individual's overall health status and more precisely predict outcomes such as the risk of cardiovascular events within 5 years. Furthermore, the additional data input into the second model (120) can be predicted as one of multiple health indicators in the first model (115) or can be input from an external source.

[0086] To explain using a specific example, the processor of the diagnostic device can input a fundus photograph of a subject into the feature extraction module (111) of the first model (115) of the risk prediction model (100). The processor can extract a feature vector of the input fundus photograph using the feature extraction module (111) and input the extracted feature vector into the output generation module (112). In addition, the processor can generate a plurality of health indices in the output generation module (112). The processor can select and extract a probability value that the CAC score is greater than 0 as a cardiovascular health indices among the plurality of health indices. In addition, the processor can input the age and gender together with the probability value that the CAC score is greater than 0 extracted from the output generation module (112) into the second model (120). At this time, the age and gender may be selected and output as health indices in the output generation module (112), or may be input from the outside. The processor can use the second model (120) to generate and output the risk of occurrence of a major cardiovascular disease event within 10 years as cardiovascular risk information.

[0087] In addition, when the processor selects and extracts a probability value that eGFR is less than 60 as a kidney health indicator among multiple health indicators, the processor may input age and gender together with the probability value that eGFR is less than 60 extracted from the output generation module (112) into the second model (120). In addition, the processor may use the second model (120) to generate a risk of occurrence of a major kidney disease event within 10 years as kidney risk information and output it.

[0088] Additionally, the processor of the diagnostic device can generate subject-specific auxiliary information based on the risk information.

[0089] The processor can generate and provide subject-tailored auxiliary information to support clinical judgment of medical staff, particularly drug prescription decisions, based on the subject's systemic disease risk information (e.g., cardiovascular risk score, renal risk score, etc.) derived from the risk prediction model (100). The subject-tailored auxiliary information can mean information processed in a form that can provide practical assistance in establishing a drug treatment strategy by going beyond simply presenting a predicted risk value and interpreting the clinical significance of the risk and evaluating it in the context of the subject's individual condition and existing risk factors.

[0090] For example, subject-specific ancillary information may include information about what a subject's risk level means in relation to treatment goal setting or drug prescribing criteria suggested by specific clinical guidelines.

[0091] Additionally, the analysis may include information on the rate of disease progression or response to treatment by analyzing the risk change trend of the subject over time, or may include analysis information on the relative importance or contribution of the risk information to the overall risk profile of the subject.

[0092] Subject-specific ancillary information can be used as support for decision-making, such as changing existing medications or adjusting the dosage. Ultimately, this subject-specific ancillary information can provide objective and personalized evidence for healthcare professionals to decide whether to initiate, adjust, change, or maintain a specific medication, such as a statin or antihypertensive, or to develop additional testing or follow-up plans.

[0093] This personalized support information may also include lifestyle modifications that involve behavioral changes in the subject. These lifestyle modifications may include, for example, dietary adjustments, improved exercise habits, smoking cessation, alcohol restriction, weight loss, or stress management. These behavioral modifications can be utilized as non-pharmacological interventions to reduce the subject's risk of cardiovascular disease. If the risk change patterns described below progress in an improved direction, it may be possible to reduce the dosage of prescribed medications and transition to these behavioral modifications.

[0094]

[0095] Additionally, the processor can generate subject-specific auxiliary information using two main methods: risk change pattern analysis or risk factor weight calculation. Furthermore, the processor can generate subject-specific auxiliary information using a hybrid method that combines risk change pattern analysis and risk factor weight calculation. Below, the risk change pattern analysis and risk factor weight calculation methods are described in detail.

[0096]

[0097] Risk change pattern analysis method

[0098] FIG. 3 is a diagram for explaining a method for analyzing a risk change pattern according to one embodiment and providing target-tailored auxiliary information using the same.

[0099] Referring to FIG. 3, a method for providing target-tailored auxiliary information according to one embodiment may include a step of obtaining risk information (S100), a step of analyzing a risk change pattern (S200), and a step of providing target-tailored auxiliary information (S300).

[0100] In step S100, the processor of the diagnostic device may acquire risk information about the subject from a storage module or an external source. Here, the risk information may include a risk score and / or a risk grade. For example, systemic disease risk information and cardiovascular risk information may be expressed as a systemic disease risk score / grade and a cardiovascular risk score / grade. Hereinafter, for convenience of explanation, risk information is expressed as a risk score, but this is not limited thereto, and the description of the risk score in this specification can of course be applied to health-related indicators, which are its superordinate concepts.

[0101] Additionally, the risk score may be a systemic disease risk score. As described above, the systemic disease risk score may include at least one of a cardiovascular risk score, a renal risk score, a liver risk score, a brain risk score, or a pulmonary and respiratory risk score. Furthermore, the systemic disease risk score may be a weighted sum of multiple detailed systemic disease risk scores. Since the aforementioned description may apply to step S100, a detailed description thereof will be omitted.

[0102] In step S200, the processor of the diagnostic device can analyze risk change patterns. Risk change pattern analysis may refer to analyzing changes in systemic disease risk over time based on risk information.

[0103] First, the processor can obtain a time-series (longitudinal) risk data set required for analyzing risk change patterns. The processor can set the risk score obtained in step S100 as the risk score at the current point in time. Furthermore, the processor can obtain past history data of the subject from a storage module or an external source. Furthermore, the processor can obtain risk scores for previous points in time, measured at previous points in time, from the past history data. Furthermore, the processor can configure the current risk score and the risk scores for previous points in time into a time-series risk data set. Furthermore, in some cases, the processor can also obtain reference data for comparison (e.g., the average risk change trajectory of a similar patient group) from the storage module or an external source. Furthermore, the processor can analyze the risk change pattern of the subject using the time-series risk data set. In one embodiment, the processor can analyze the change trend of the risk score over time as the risk change pattern of the subject based on the concepts of “speed” and / or “acceleration.”

[0104] More specifically, the processor can detect a risk acceleration indicator based on the acquired time-series risk data set as an analysis of risk change patterns. The risk acceleration indicator can also be expressed as an Accelerator indicator. The risk acceleration indicator is based on the concept of acceleration of a moving object and can refer to a point where the slope of the target's risk score change pattern changes abruptly. First, the meaning of the risk acceleration indicator will be explained using Figure 4.

[0105]

[0106] Figure 4 is an exemplary graph illustrating the meaning of a risk acceleration index according to one embodiment. In Figure 4, the horizontal axis represents time (e.g., the number of years since diabetes diagnosis), and the vertical axis represents a risk score.

[0107] Referring to Fig. 4, the time-series risk data set of the subject can be expressed as a biomarker score according to each measurement time. Here, the biomarker score can include a systemic disease risk score such as a cardiovascular risk score or a renal risk score. In the graph, line (1001) is an 'average trajectory', which can correspond to a reference line showing the average risk change of a specific patient group. Line (1002) is a 'patient trajectory with intervention', or a 'target risk trajectory', which can represent an ideal risk change trajectory that can be expected when drug treatment or lifestyle management is performed well. Line (1003) is a 'patient trajectory without intervention', which shows a pattern in which the risk gradually increases steeply due to improper risk management.

[0108] As mentioned above, the risk acceleration index is the point at which a sharp change in the risk score occurs, which can be observed in the area indicated by the rectangle (1004) on the graph (e.g., at approximately 2.37 years of disease duration). Such a sharp change in the risk score can be an important signal indicating a change in the patient's health status or the need for medication intervention. For example, in the patient trajectory without medication intervention (line (1003)), the slope of the risk acceleration index increases sharply after the detection point, indicating a trend of rapid increase in the risk score. This may indicate that the risk acceleration index can signal the onset or worsening of complications. In contrast, in the patient trajectory with medication intervention (line (1002)), the slope becomes flatter at the same point, indicating a suppression of the increase in the risk score. This may suggest that medical intervention at a point at which complications may develop or worsen can effectively control disease progression.

[0109]

[0110] Returning to the description of step S200 of FIG. 3, the processor can detect a risk acceleration indicator through the following process. Hereinafter, for convenience of explanation, the risk acceleration indicator is also expressed as an acceleration indicator.

[0111] First, the processor can preprocess time-series risk data. It collects and normalizes time-series risk data for a target object, filtering out outliers and noise that may exist in the time-series data to improve analysis accuracy. For data collected at irregular intervals, the time intervals can also be adjusted to a consistent level.

[0112] Second, the processor can calculate the slope of the risk score. The processor can calculate the slope by calculating the change in the risk score between two consecutive measurement points. The slope can be calculated as (change in risk score) ÷ (time interval). For example, if the risk score at time t1 is 25 and at time t2 is 30, and the time interval is 0.5 years, the slope would be (30-25) ÷ 0.5 = 10.

[0113] Third, the processor can detect an acceleration index. Since the acceleration index is the amount of increase in the slope, the processor can calculate the change in the slope of consecutive intervals. The acceleration index at a specific time point (t) can be calculated as the change in the slope between time points t to t+1 and the slope between t-1 and t. For example, if the acceleration index is evaluated at time point 2.37, this time point corresponds to between years 2 and 3. If the slope of the previous interval, that is, the slope of the 1-2 year interval, is 8, and the slope of the subsequent interval, that is, the slope of the 2-3 year interval, is 14, the acceleration index is 6. This patient's risk score increased at an accelerated rate at year 2.37, and the acceleration index is 6.

[0114] Fourth, the processor can perform statistical significance verification. In one embodiment, the processor can evaluate the appropriateness of the current risk change rate by comparing the slope change rate in the acceleration index of the calculated subject with at least one of 1) the subject's past change rate pattern, 2) the average trajectory (e.g., line (1001) of FIG. 4) which represents a reference line showing the average risk change of a specific patient group, or 3) the target risk change trajectory (e.g., line (1002) of FIG. 4) that can be expected when drug treatment or lifestyle management is performed well. If the slope change rate in the acceleration index is higher than a threshold when compared with at least one of the subject's past change rate pattern, the average trajectory change rate, or the target risk change trajectory change rate, the processor can determine that the acceleration index is statistically significant. Furthermore, in one embodiment, the processor can utilize a change point detection algorithm to verify the statistical significance of the detected acceleration index. Points where statistically significant changes occur in risk time series data can be identified through algorithms such as CUSUM (Cumulative Sum), PELT (Pruned Exact Linear Time), and Bayesian change point detection.

[0115] For example, a cohort of patients could be tracked for cardiovascular risk over a five-year period, and an acceleration index could be calculated at specific time points (e.g., 2 to 3 years). Subjects could then be divided into three groups based on their acceleration index values, as follows:

[0116] 1) Low-risk group: acceleration index ≤ 3

[0117] 2) Medium-risk group: 3 < acceleration index ≤ 6

[0118] 3) High-risk group: Acceleration index > 6

[0119] In this case, since the risk of developing complications is considered to significantly increase when the acceleration index is 6 or higher, the processor can set the threshold (cutoff) to 6. In the example above, the subject in question can be evaluated as belonging to the borderline between the medium-risk group and the high-risk group since the acceleration index is 6 based on 2.37 years.

[0120] Additionally, the processor can analyze the clinical significance of the risk acceleration index. Analyzing the clinical significance of the risk acceleration index goes beyond simply noting that the rate of risk increase has accelerated, and can include interpreting its significance within a clinical context. This can be based on the understanding in the present invention that rapid changes in the risk score can signal the development of future complications or worsening of the condition.

[0121] Additionally, the processor can compare the acceleration index and complication rate for each risk group. For example, the five-year complication rate can be 8% for the low-risk group, 19% for the medium-risk group, and 36% for the high-risk group. A patient with an acceleration index of 6 or higher is at a high risk of developing complications, which can be used as a basis for determining the need for early intervention.

[0122]

[0123] Figure 5 is a diagram for explaining the distribution of risk scores according to the duration of diabetes and the presence or absence of complications according to one embodiment.

[0124] Referring to Figure 5, the distribution of cardiovascular disease risk scores according to diabetes duration (0-5 years, 5-10 years, 10 years or more) and the presence or absence of complications (No Cx indicates no complications, Cx indicates complications) is represented in a box plot. The p value above each box indicates the statistical significance of the difference in scores according to the presence or absence of complications within the corresponding diabetes duration group.

[0125] In Fig. 5, the following clinical implications can be confirmed.

[0126] 1) In the group with a disease duration of 0-5 years, the risk score of patients with complications (Cx) was statistically significantly higher than that of patients without complications (No Cx) (p=0.016). This suggests that the risk score increases significantly when complications develop in patients with early-onset diabetes.

[0127] 2) Although the risk score of patients with complications tended to be higher in the group with a disease duration of 5-10 years, the difference may have a slightly lower statistical significance than that of the early group (p=0.072).

[0128] 3) In the group with a disease duration of 10 years or more, the difference in risk scores based on the presence or absence of complications was not statistically significant (p=0.49). This suggests that risk scores tend to increase overall in patients with long-term diabetes, regardless of the presence or absence of complications.

[0129] These analytical results can provide valuable background for analyzing the clinical significance of cardiovascular risk scores (or grades) as complications of diabetes. For example, a high cardiovascular risk score in patients with early diabetes (0-5 years) can be interpreted as a clinically significant signal, indicating a higher risk of developing complications.

[0130] In Figure 5, the correlation between cardiovascular risk scores and complications in diabetic patients is considered in conjunction with the acceleration indicator according to the present invention, providing useful supplementary information. Specifically, the information is as follows.

[0131] Returning to step S200, the processor may acquire subject characteristic information from a storage module or an external source. The subject characteristic information may include the subject's age, sex, body mass index (BMI), disease duration, presence or absence of complications, medication prescription history, patient images, etc. In addition, the processor may analyze the clinical significance of the risk acceleration indicator based on the subject characteristic information. For example, as described above, the processor may verify the clinical significance of the risk acceleration indicator based on the subject's disease duration and presence or absence of complications. Referring to the contents described in FIG. 5, when the subject's diabetes duration is 5 years or less and there are no diabetic complications, and the subject's risk acceleration indicator is detected, the processor may determine that the subject is likely to develop diabetic complications.

[0132] Additionally, the processor can perform a treatment responsiveness analysis of the risk acceleration indicator as one of the analyses of the clinical significance of the risk acceleration indicator. This includes a process of predicting the expected treatment effect when a treatment intervention (e.g., a change in medication prescription) is performed on a subject in whom the risk acceleration indicator is detected. As illustrated in FIG. 4, the processor can predict the potential effect of the treatment intervention on the risk acceleration indicator detected in the subject by comparing the trajectory without the medication prescription intervention (line (1003)) and the trajectory with the medication prescription intervention (line (1002)). In this case, reference data such as that shown in FIG. 4 can be stored in advance in the storage module, and the processor can obtain the reference data from the storage module and use clinical data to predict the potential effect of the treatment intervention on the risk acceleration indicator detected in the subject.

[0133] Furthermore, the processor can synthesize all analysis results to determine the clinical significance of the final risk acceleration indicator and utilize this information to generate subject-specific supplementary information. In this way, the processor can comprehensively consider the magnitude of the risk acceleration indicator, the timing of detection, clinical significance, subject characteristics, and expected treatment response to derive supplementary information optimized for the subject.

[0134]

[0135] Additionally, in step S300, the processor of the diagnostic device may generate and provide subject-specific auxiliary information based on the results of the risk change pattern analysis. The generated subject-specific auxiliary information may be determined by factors such as the subject's condition, disease duration, whether an accelerated risk indicator is detected, and clinical significance.

[0136] In one embodiment, the processor can generate different types of auxiliary information depending on whether a risk acceleration indicator is detected and its magnitude. For example, if a risk acceleration indicator is detected above a threshold, auxiliary information such as "The risk is rapidly increasing, so you may consider strengthening your medication regimen" may be provided. Conversely, if a risk acceleration indicator is below the threshold or not detected, auxiliary information such as "The risk increase is being managed stably, so you may maintain your current medication regimen" may be provided.

[0137] Additionally, the processor can generate customized supplementary information based on the duration of the subject's disease. For subjects with early diabetes (duration of disease 0-5 years), accelerated risk indicators may be a precursor to the development of complications, so supplementary information such as, "This subject is a subject with early diabetes. Accelerated risk indicators have been detected, indicating an increased risk of developing complications. Consider early therapeutic intervention." Conversely, for subjects with long-term diabetes (duration of disease 10 years or more), supplementary information such as, "This subject is a subject with long-term diabetes. Accelerated risk indicator detection may be part of the natural progression of the disease. Management through regular monitoring is important."

[0138] Additionally, the processor can provide information about expected treatment response as part of the subject-specific ancillary information. For example, it can provide specific information related to drug prescriptions, such as, "The subject's risk profile is expected to respond well to drug treatment. Consider prescribing a statin."

[0139] Additionally, the processor can provide information comparing the subject's risk score to the average risk of a reference group as part of the subject-specific supplementary information. For example, by providing relative positional information such as "the subject's risk score is 20% higher than the average risk of a group of subjects with similar disease duration and comorbidities," this can help medical professionals objectively assess the subject's condition.

[0140] Additionally, the processor can provide personalized auxiliary information to the subject in various forms, such as text, graphs, and tables. In particular, visualizations, such as those shown in Figure 4, can intuitively convey the location, size, and clinical significance of risk acceleration indicators, thereby aiding medical staff's understanding.

[0141] The patient-specific supplementary information generated in this way can serve as valuable reference for medical professionals when making drug prescribing decisions, particularly in determining when to start a medication, whether to adjust dosage, or whether to change medication. Furthermore, it can contribute to providing precise medical services tailored to the individual patient's disease progression pattern.

[0142] The method for providing personalized supplementary information described above goes beyond simply providing a current risk score. By analyzing risk changes over time, particularly risk acceleration indicators, it can capture the dynamics of disease progression and interpret this within a clinical context, providing more useful information to medical professionals. This allows medical professionals to make personalized medication decisions that take into account the individual patient's rate and pattern of disease progression, potentially enabling more effective disease management.

[0143]

[0144] Risk factor weight calculation method

[0145]

[0146] FIG. 6 is a diagram for explaining a method for calculating risk factor weights according to one embodiment and a method for providing target-tailored auxiliary information using the same.

[0147] Referring to FIG. 6, a method for providing target-tailored auxiliary information according to one embodiment may include a step of obtaining risk information (S100), a step of calculating a risk factor weight (S400), and a step of providing target-tailored auxiliary information (S500).

[0148] In step S100, the processor of the diagnostic device may obtain cardiovascular risk information for the subject from a storage module or an external source. Furthermore, the aforementioned content may also apply to step S100, and therefore, a detailed description thereof will be omitted.

[0149] Cardiovascular risk scores may have technological significance as novel biomarkers, comparable to traditional risk factors (age, smoking, obesity, hypertension, diabetes, dyslipidemia, family history, etc.). These novel biomarkers, extracted from medical data such as fundus images, can provide a noninvasive alternative for assessing cardiovascular disease risk without the need for conventional invasive blood tests or physical examinations. This not only increases patient convenience, but also improves healthcare accessibility and reduces testing costs. This will be described in detail using Figure 7.

[0150]

[0151] FIG. 7 is a diagram illustrating the usefulness of a cardiovascular risk score as a new biomarker according to one embodiment.

[0152] Referring to Fig. 7, (A) is a graph showing the cumulative cardiovascular disease incidence rate of the low-risk group and the high-risk group over time according to the cardiovascular risk score, and (B) is a graph showing the cumulative cardiovascular disease incidence rate of the low-risk group and the high-risk group over time according to the existing 10-year AtheroSclerotic CardioVascular Disease (ASCVD) risk. In each graph, the horizontal axis represents time (years), and the vertical axis represents the cumulative cardiovascular disease incidence rate (Cumulative events, %). The cardiovascular risk score referred to here means the result of predicting the risk of cardiovascular disease incidence within a given period by learning the correlation with the coronary artery calcification index (CAC) by the risk prediction model (100) of Fig. 2 (A) and (B) based on the fundus image.

[0153] Comparing the two graphs, we can see that the pattern of cardiovascular disease incidence over time is almost identical. That is, the high-risk group with a cardiovascular risk score of (A) (line (1011)) and the high-risk group with a 10-year ASCVD risk of 10% or more (line (1013)) in (B) both show a steadily increasing incidence rate during the follow-up period, reaching approximately 6% at 10 years. In contrast, the low-risk group with a cardiovascular risk score of (A) (line (1012)) and the low-risk group with a 10-year ASCVD risk of less than 10% (line (1014)) maintain relatively low incidence rates, remaining at approximately 2% at 10 years.

[0154] Additionally, Table 1 below can statistically support this visual similarity.

[0155]

[0156] Number of biomarkers (N) Number of cases (Cases) Person-years (Person-Years) Incidence (95% CI) Hazard ratio (95% CI) Cardiovascular risk score_Low risk (Low) 7680129763241.68 (1.68, 1.68) 1 (reference) Cardiovascular risk score_High risk (High) 1817120176866.60 (6.59, 6.62) 3.94 (3.13, 5.00) C-statistic (C-stat) 0.647 (0.617, 0.678) ASCVD risk < 10% 7680131763581.71 (1.70, 1.71) 1 (reference) ASCVD risk ≥ 10%1817118176526.49 (6.48, 6.51)3.65 (2.73, 4.84)C-statistic(C-stat)0.644(0.613, 0.675)

[0157] Referring to Table 1, the hazard ratio (Hazard Ratio) of the high-risk group according to the cardiovascular risk score was 3.94 (95% CI: 3.13-5.00), indicating a risk of developing cardiovascular disease that was approximately four times higher than that of the low-risk group. Very similarly, the Hazard Ratio of the high-risk group with a 10-year ASCVD risk of 10% or more was 3.65 (95% CI: 2.73-4.84). More importantly, the concordance index (C-stat), which objectively evaluates the predictive performance of the two models, was 0.647 (95% CI: 0.617-0.678) for the cardiovascular risk score and 0.644 (95% CI: 0.613-0.675) for the ASCVD risk, showing almost statistically identical levels. This may mean that medical data-based cardiovascular risk scores have predictive accuracy equivalent to ASCVD risk scores derived from a combination of complex blood tests and clinical information.

[0158] This could mean that a cardiovascular risk score based on medical data has predictive accuracy equivalent to a 10-year ASCVD risk score calculated by synthesizing complex blood tests and clinical information. In particular, a risk score cutoff corresponding to the 10% ASCVD risk threshold, which is used as an important criterion for determining whether to receive drug treatment in stage 1 hypertensive patients, can be set using the cardiovascular disease risk score according to the present invention, suggesting that the score can be directly utilized as a drug prescribing criterion in actual clinical practice.

[0159] Furthermore, these results demonstrate that the cardiovascular risk score not only has predictive performance similar to that of the existing 10-year ASCVD composite score, but also has sufficient clinical value as an independent predictor variable when compared with individual traditional risk factors.

[0160] In addition, the cardiovascular risk grade classification according to one embodiment of the present invention has statistical significance as an independent predictive variable compared to existing traditional cardiovascular risk factors. Specifically, according to the results of univariable analysis in Table 2, the medium-risk group of the cardiovascular risk grade (the cardiovascular risk score presented in the present embodiment means the cardiovascular risk grade predicted by AI that learned the correlation between the fundus image and the CAC score) showed OR 1.898 (95% confidence interval: 1.828-1.971), and the high-risk group showed OR 3.494 (95% confidence interval: 3.406-3.585), which indicates a high risk level similar to existing risk factors such as hypertension (OR 4.989) and diabetes (OR 3.595).

[0161] Furthermore, even after adjusting for existing risk factors such as age, smoking, hypertension, and diabetes through multivariable analysis, the intermediate-risk group of the cardiovascular risk grade maintained statistically significant results with OR 1.352 (95% confidence interval: 1.296-1.410) and the high-risk group with OR 2.148 (95% confidence interval: 2.081-2.218).

[0162] Therefore, the cardiovascular risk rating of the present invention can function as an independent risk factor with predictive power equal to or similar to that of existing traditional risk factors, beyond a simple auxiliary indicator, suggesting that it can be utilized as a key judgment indicator when setting treatment or drug prescription criteria in the future.

[0163]

[0164] VariablesUnivariate Analysis (Odds Ratio, OR: 95% CI)Multivariate Analysis (Odds Ratio, OR: 95% CI)Age3.096 (3.005 - 3.190)1.689 (1.591 - 1.689)Smoking1.174 (1.131 - 1.219)1.337 (1.283 - 1.392)Hypertension4.989 (4.863 - 5.118)3.990 (3.885 - 4.098)Diabetes3.595 (3.419 - 3.780)2.046 (1.936 - 2.161)Cardiovascular Risk Score_Middle Risk Group1.898 (1.828 - 1.971)1.352 (1.296 - 1.410)Cardiovascular Risk Score_High Risk Group 3.494 (3.406 - 3.585)2.148 (2.081 - 2.218)

[0165] Meanwhile, in the high-risk group with a cardiovascular risk score, statin administration was shown to have the effect of reducing the risk of cardiovascular disease by 36%, as shown in Table 3.

[0166]

[0167] Overall high-risk group Hazard ratio (HR, 95% CI) Hazard ratio (HR, 95% CI) Statin non-user 1.00 (Reference) 1.00 (Reference) Statin user 0.97 (0.72 - 1.30) 0.64 (0.42 - 0.97)

[0168] Specifically, Table 3 above shows that among patients identified as high-risk by AI that learned the correlation between fundus images and CAC scores, statin users (n=896) had a significantly lower risk of cardiovascular events than non-users (n=896). The hazard ratio (HR) was 0.64 (95% confidence interval: 0.42–0.97), indicating a 36% risk reduction with statin use. In contrast, the hazard ratio in the overall patient group before matching was 0.97 (95% CI: 0.72–1.30), which was not statistically significant. These results suggest that AI that learned the correlation between fundus images and CAC scores can effectively select patients who are likely to benefit from statin treatment.

[0169] This suggests that cardiovascular risk scores are valuable not only for predicting risk but also as indicators of actual treatment effectiveness. In other words, this provides strong evidence that drug prescribing decisions based on cardiovascular risk scores can actually lead to improved clinical outcomes.

[0170] Thus, cardiovascular risk scores may have value as novel biomarkers equivalent to traditional cardiovascular risk factors in the following aspects:

[0171] (1) Predictive performance: Predictive performance based on cardiovascular risk score shows predictive accuracy equivalent to that of existing ASCVD risk assessment models.

[0172] (2) Independent predictive power: The cardiovascular risk score maintains independent predictive value even when considered together with traditional risk factors.

[0173] (3) Prediction of treatment effect: The risk reduction effect through statin treatment can be predicted using the cardiovascular risk score.

[0174] (4) Non-invasive: Cardiovascular risk can be assessed using only medical data such as fundus images, without blood tests.

[0175] (5) Drug prescribing criteria: Cardiovascular risk scores can be directly used to make specific drug prescribing decisions.

[0176]

[0177] Risk Change Pattern Analysis Method: Calculating Weights Between Cardiovascular Risk Grades and Existing Cardiovascular Risk Factors

[0178] Returning to Figure 6, at step S400, the processor can calculate risk factor weights. Specifically, the processor can calculate relative weights between the acquired cardiovascular risk information and existing cardiovascular risk factors. This step can quantitatively assess how the cardiovascular risk score, along with other existing risk factors, contributes to the subject's cardiovascular disease risk, thereby supporting personalized medication prescribing decisions.

[0179] In one embodiment, to calculate risk factor weights, the processor may apply an attention method. Attention is a technology developed in the field of deep learning that enables data-driven learning of which input features have a greater impact on outcome prediction. This has the advantage of dynamically determining the importance of each risk factor based on the characteristics of each individual subject, rather than simply assigning equal weights to all risk factors. The application of attention is illustrated in Figure 8.

[0180] Figure 8 is a diagram for explaining an attention weight distribution according to one embodiment.

[0181] Referring to FIG. 8, the processor can process a dataset including input variables of glycated hemoglobin (HBA1c), age, gender, systolic blood pressure (SBP), diastolic blood pressure (DBP), and CAC predicted value. Here, the CAC predicted value represents a predicted coronary artery calcification score based on a retinal image, and the CAC predicted value may indicate a cardiovascular disease risk. The CAC predicted value may be an example of the aforementioned cardiovascular risk score. All of these features are input to the model in a normalized form, and the model can be trained to predict correct labels related to cardiovascular disease factors based on them.

[0182] At this point, the processor calculates the contribution of each input variable in the form of a learnable attention score and, through softmax normalization, calculates a relative attention weight between 0 and 1. The calculated weights can vary by subject, and key predictors can dynamically change based on the clinical characteristics of each subject.

[0183] For example, the case illustrated in (A) of Figure 8 shows the attention weight distribution for patients with high HbA1c levels (i.e., the presence or absence of HbA1c >0.2 was used as the correct label), and in this case, the weight of HbA1c was the highest at 0.25, followed by age (0.14), gender (0.13), systolic blood pressure (0.13), diastolic blood pressure (0.13), and CAC prediction value (0.12). This suggests that HbA1c is the most important factor in predicting cardiovascular risk in this case.

[0184] On the other hand, in the case shown in (B) of Figure 8, for patients with high CAC prediction values ​​(i.e., the presence or absence of CAC >0.2 is used as the correct label), the weight of gender is the highest at 0.24, and the CAC prediction value is calculated as 0.20 and the glycated hemoglobin as 0.20, indicating that gender, glycated hemoglobin, and the retina-based CAC prediction value act as major predictive factors.

[0185] Through this embodiment, the processor can automatically extract key risk factors for each subject, which can be utilized for patient-tailored decision-making or drug prescription.

[0186] Meanwhile, in addition to attention methods, the processor can apply techniques such as SHapley Additive ExPlanations (SHAP), permutation importance, or logistic regression-based coefficient analysis to calculate the contribution of each risk factor. Compared to attention methods, these methods offer the advantage of increasing model interpretability and maintaining a simple structure.

[0187] The processor can store the results of these attention-based weight calculations in the storage module and utilize them to generate subject-specific auxiliary information in step S500 of FIG. 6. This allows medical professionals to identify the factors that have the greatest impact on the subject's cardiovascular disease risk and use this as an objective basis for determining whether to prescribe medication and its intensity. For example, in the case of a subject whose cardiovascular risk score weights significantly higher than other risk factors, medical professionals can consider prescribing a statin more aggressively than existing drug prescribing guidelines based on this. Conversely, if the weights of other risk factors are high and the weight of the cardiovascular risk score is relatively low, a treatment strategy can be established that focuses more on managing that risk factor.

[0188]

[0189] Returning to FIG. 6 again, in step S500, the processor may also generate and provide subject-specific auxiliary information based on the risk factor weights calculated in step S400.

[0190] For example, the processor may generate customized auxiliary information, including at least one of the following: drug prescription criteria, drug prescription strength, drug prescription type, whether to change the drug prescription, or drug prescription start time, based on the risk factor weights calculated in step S400. Through step S500, the processor may provide customized information that medical professionals can use as an objective basis when making decisions about prescribing drugs to prevent cardiovascular disease.

[0191] In one embodiment, the processor can analyze the risk factor weights calculated using attention methods to generate supplementary information that emphasizes the relative importance of the cardiovascular risk score, particularly when this score is found to be highly significant. For example, a recommendation could be provided such as, "The cardiovascular risk score weights were calculated to be high compared to other risk factors, so it is recommended to adjust the medication prescription intensity accordingly."

[0192] Specific examples of drug prescriptions include information on statin prescriptions and blood pressure-lowering medication prescriptions. Table 2 below provides examples of subject-specific supplementary information for statin prescriptions:

[0193]

[0194] Subject LDL-C value Existing statin prescription criteria Cardiovascular risk score Subject-tailored supplementary information Subject A 140 mg / dL Prescription not recommended High-risk statin prescription possible Subject B 180 mg / dL Prescription consideration Low-risk statin prescription possible relaxation

[0195] As shown in Table 4, for subject A, if the LDL-C level is 140 mg / dL, which is below the general statin prescription standard, but the cardiovascular risk score is also high, the processor can provide subject-tailored supplementary information such as, “The LDL-C level is lower than the prescription standard, but the cardiovascular risk score is high, so we recommend considering statin prescription.”

[0196] Conversely, for subject B, if the LDL-C level is 180 mg / dL, which exceeds the typical statin prescription criteria, but the cardiovascular risk score is low, the processor can provide supplementary information such as, "Although the LDL-C level is high, the cardiovascular risk score is low, so you may reconsider prescribing a statin or consider prescribing a lower dose."

[0197] In this way, the processor can help more precisely adjust statin prescriptions by considering not only traditional cardiovascular risk factors like LDL-C levels but also cardiovascular risk scores. This could provide research opportunities to improve existing guidelines.

[0198] A similar approach can be applied to prescribing antihypertensives. For example, in patients with stage 1 hypertension (systolic blood pressure 130-139 mmHg or diastolic blood pressure 80-89 mmHg), initiating drug treatment was previously recommended if the 10-year ASCVD risk was 10% or higher. However, as shown in Figure 7, a cardiovascular risk score can have predictive power equivalent to ASCVD risk. Accordingly, the processor can provide customized supplementary information on whether to prescribe antihypertensives based on the cardiovascular risk score. Table 3 presents specific examples of antihypertensive drug prescribing.

[0199]

[0200] Subject blood pressure values: Existing blood pressure lowering medication prescription criteria; Cardiovascular risk score; Subject-tailored supplementary information; Subject C: 135 / 85 mmHg; Lifestyle modification is recommended; High-risk: Early prescription of blood pressure lowering medication is recommended; Subject D: 145 / 90 mmHg; Drug therapy is recommended; Low-risk: Reevaluation possible after attempting lifestyle modification

[0201] In Table 5, subject C has stage 1 hypertension (systolic blood pressure 135 mmHg, diastolic blood pressure 85 mmHg). According to existing guidelines, lifestyle modifications are generally considered a priority over drug treatment. However, because the patient's cardiovascular risk is high, the processor can provide tailored supplementary information, such as, "Although blood pressure is borderline, the cardiovascular risk score is very high. Therefore, early prescribing of antihypertensives along with lifestyle modifications is recommended."

[0202] In contrast, subject D has stage 2 hypertension (systolic blood pressure 145 mmHg, diastolic blood pressure 90 mmHg), a condition for which standard guidelines recommend immediate drug treatment. However, he also presents a low cardiovascular risk. In this case, the processor could provide subject-specific supplementary information, such as, "While his blood pressure meets the criteria for drug treatment, his cardiovascular risk score is low. Therefore, intensive lifestyle modifications may be considered first, followed by reevaluation after 1-3 months."

[0203] As an additional example, a case in which blood pressure and other risk factors coexist may be considered. Table 4 presents a specific example that considers risk factors in combination.

[0204]

[0205] Subject Blood pressure value Existing risk factors Cardiovascular risk score Subject-specific supplementary information Subject E 138 / 88 mmHg Diabetes (O), smoking (O) High risk Active drug treatment recommended (consider polypharmacy) Subject F 142 / 92 mmHg Diabetes (X), smoking (X) Low risk Start with a low dose of a single drug, gradually adjust

[0206] For subject E, he has stage 1 hypertension, diabetes, and smoking, as well as a high cardiovascular risk. In this case, the processor could provide supplementary information such as, "Considering the combined risk factors and high cardiovascular risk score, immediate and aggressive drug treatment is recommended. If necessary, combination therapy with two classes of antihypertensives can be considered from the outset."

[0207] Subject F has stage 2 hypertension, but no other major risk factors and an intermediate cardiovascular risk. In this case, the processor could provide supplementary information such as, "Blood pressure meets the criteria for drug treatment, and drug treatment is indicated given the moderate cardiovascular risk score. However, starting with a low dose of a single drug and observing the response while gradually titrating is recommended."

[0208] This personalized supplementary information can help clinicians make more precise treatment decisions by considering overall cardiovascular risk, not just blood pressure readings. In particular, by integrating novel biomarkers, such as cardiovascular risk scores, with traditional risk factors, it can help develop optimized drug prescribing strategies tailored to the individual patient's characteristics.

[0209] Additionally, the processor can provide personalized supplementary information with various designs and interfaces. For example, personalized supplementary information can include text-based prescription recommendations along with visual information, such as a chart. For example, the chart can include six major risk factors: smoking, obesity, hypertension, diabetes, dyslipidemia, and a risk score. In the chart, the risk score represents the cardiovascular risk score, while the remaining factors correspond to traditional cardiovascular risk factors. The importance of each risk factor is divided into three levels: low, medium, and high, which can help medical professionals easily understand the relative importance of each risk factor. Such a chart can intuitively provide medical professionals with the following important information: First, they can easily identify at a glance which risk factor has the greatest impact on the subject's cardiovascular risk. The chart above shows that cardiovascular risk scores and obesity have the highest weights, suggesting that these two factors contribute most significantly to the subject's cardiovascular disease risk. Furthermore, the chart can be used to compare the relative importance of risk factors. Furthermore, the chart can assist in prioritizing treatment strategies. The processor can provide information indicating that a treatment approach focused on high-weighted risk factors may be effective.

[0210] Along with these charts, the processor can provide additional textual explanations or recommendations. For example, the processor could provide customized supplementary information such as, "The subject's cardiovascular risk score and obesity weighting were calculated to be higher than other risk factors, so we recommend that you plan medication prescriptions and lifestyle modifications accordingly. In particular, considering the high weighting of the cardiovascular risk score, aggressive medication may be considered even if current LDL-C levels or blood pressure are at borderline levels." These visualization tools can be effective in supporting evidence-based, personalized treatment decisions by conveying complex risk factor weighting information to healthcare professionals in an intuitive and understandable format. Furthermore, healthcare professionals can utilize these visual tools to more effectively explain risk factors and the need for treatment when discussing treatment plans with patients.

[0211]

[0212] FIG. 9 is a drawing for explaining a method for providing target-tailored auxiliary information according to one embodiment.

[0213] Referring to FIG. 9, in step S1000, the processor of the diagnostic device can input the medical data of the subject into a risk prediction model to process the medical data.

[0214] Additionally, in step S2000, the processor of the diagnostic device can obtain risk information of the subject as a result of processing medical data from the risk prediction model.

[0215] Additionally, in step S3000, the processor of the diagnostic device can provide customized auxiliary information for the subject based on the risk information.

[0216] For steps S1000 to S3000, the above-mentioned matters may be applied, so a detailed description thereof is omitted.

[0217]

[0218] Various embodiments of the present specification may be implemented as software including instructions stored on a machine-readable storage medium that can be read by a machine (e.g., a computer). The machine is a device that can call instructions stored from the storage medium and operate according to the called instructions, and may include an electronic device according to the disclosed embodiments. When the instructions are executed by a processor, the processor can perform a function corresponding to the instructions directly or under the control of the processor using other components. The instructions may include code generated or executed by a compiler or interpreter. The machine-readable storage medium may be provided in the form of a non-transitory storage medium. Here, the term "non-transitory storage medium" does not mean that it does not contain signals and is tangible, but does not distinguish whether data is stored semi-permanently or temporarily in the storage medium. For example, a "non-transitory storage medium" may include a buffer in which data is temporarily stored.

[0219] According to one embodiment, the method according to the various embodiments disclosed in the present specification may be provided as a computer program product. The computer program product may be traded as a product between a seller and a buyer. The computer program product may be distributed in the form of a device-readable storage medium (e.g., Compact Disc Read Only Memory, CD-ROM) or online through an application store (e.g., Play Store™). In the case of online distribution, at least a portion of the computer program product, e.g., a downloadable app, may be temporarily stored or temporarily generated in a storage medium such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0220] Although the embodiments described above have been described by way of limited examples and drawings, those skilled in the art will appreciate that various modifications and variations can be made based on the above teachings. For example, appropriate results can still be achieved even if the described techniques are performed in a different order than described, and / or components of the described systems, structures, devices, circuits, etc. are combined or combined in a different manner than described, or are replaced or substituted with other components or equivalents.

[0221] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims described below.

Claims

1. In a device for providing customized auxiliary information to a target, Contains at least one processor, At least one processor, The medical data of the subject is input into a systemic disease risk prediction model to process the medical data, Obtaining risk information of the subject as a result of processing the medical data from the above systemic disease-related risk prediction model, Provide customized auxiliary information for the subject based on the above risk information, Generating the customized auxiliary information by using at least one of a risk change pattern analysis method for analyzing changes in the risk of systemic diseases over time based on the risk information or a risk factor weight calculation method for calculating a weight of the risk information as one of multiple risk factors for cardiovascular disease. device.

2. In paragraph 1, The above systemic diseases are, Including at least one of cardiovascular disease, renal disease, liver disease, neurological disease, diabetes, or hypertension, device.

3. In paragraph 1, When generating the customized auxiliary information using the above risk change pattern analysis method, At least one processor, Obtain a time series risk data set consisting of the risk information obtained above and the risk information of a previous point in time, By analyzing the above time series risk data set, risk acceleration indicators are detected, By analyzing the clinical significance of the above risk acceleration indicators, the customized auxiliary information is generated. device.

4. In paragraph 3, At least one processor, In the above time series risk data set, the point where the slope change rate of risk information over time exceeds a predefined threshold is selected as a risk acceleration indicator, To verify the statistical significance of the above risk acceleration indicator, device.

5. In paragraph 1, When generating the customized auxiliary information using the above risk factor weight calculation method, The above risk information is cardiovascular risk information, At least one processor, Calculate the relative weights between the above risk information and multiple cardiovascular risk factors, Generating the customized auxiliary information based on the above relative weights, device.

6. In paragraph 4, At least one processor, Generating drug prescription information for the subject as customized auxiliary information based on the above risk acceleration index, device.

7. In paragraph 5, The above multiple cardiovascular risk factors include at least one of age, smoking, obesity, hypertension, diabetes, dyslipidemia, and family history. device.

8. In paragraph 1, The above medical data is, At least one of ophthalmology-related medical data, cardiovascular-related medical data, kidney-related medical data, liver-related medical data, brain-related medical data, or lung and respiratory-related medical data, device.

9. In paragraph 5, At least one processor, Generating drug prescription information for the subject as customized auxiliary information based on the risk information and the relative weights between multiple cardiovascular risk factors. device.

10. In a method for providing customized auxiliary information of a target object, A step of inputting the medical data of the subject into a systemic disease risk prediction model and processing the medical data; A step of obtaining risk information of the subject as a result of processing the medical data from the above systemic disease-related risk prediction model; Including a step of providing customized auxiliary information for the subject based on the above risk information, The step of providing customized auxiliary information for the subject based on the above risk information is: Generating the customized auxiliary information by using at least one of a risk change pattern analysis method for analyzing changes in the risk of systemic diseases over time based on the risk information or a risk factor weight calculation method for calculating a weight of the risk information as one of multiple risk factors for cardiovascular disease. method.

11. A recording medium having recorded thereon a computer program for performing the method of Article 10.

Images