Artificial intelligence-based diabetic foot disease prediction device and method

An AI-based method using a multimodal deep learning model with foot radiographic images and electronic medical records addresses the limitations of traditional diabetic neuropathy detection, offering early and accurate predictions of diabetic foot disease.

WO2026235408A1PCT designated stage Publication Date: 2026-11-12AJOU UNIV IND ACADEMIC COOP FOUND +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2026/095739
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2026-01-29
Filing Date
2026-04-09
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Traditional diagnostic methods for diabetic peripheral neuropathy are subjective, costly, and lack sensitivity, making early detection difficult, and existing technologies like Nerve Conduction Studies are not accessible in all medical settings.

Method used

An artificial intelligence-based method using a multimodal deep learning model that combines foot radiographic images and electronic medical record data to predict diabetic peripheral neuropathy, employing a ResNet50 structure and integrating image and clinical features for early detection.

Benefits of technology

Enables early and accurate prediction of diabetic foot disease, providing cost-effective and accessible tools for resource-constrained environments, with improved generalization performance and reduced overfitting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure KR2026095739_12112026_PF_FP_ABST
    Figure KR2026095739_12112026_PF_FP_ABST
Patent Text Reader

Abstract

An artificial intelligence-based diabetic foot disease prediction device and method are disclosed. The artificial intelligence-based diabetic foot disease prediction method according to one embodiment of the present application may comprise the steps of: (a) collecting training data including image data of a foot region and electronic medical record data, which are acquired from a plurality of patients; (b) extracting image features from the image data of the training data, extracting clinical features from the electronic medical record data of the training data, and constructing a multimodal data set in which the image features and the clinical features are combined; (c) training an artificial intelligence-based disease prediction model by using the multimodal data set; (d) acquiring target data including image data and electronic medical record data of a subject to be analyzed; and (e) inputting the target data into the disease prediction model so as to output prediction information including the likelihood of onset of diabetic foot disease in the subject to be analyzed.
Need to check novelty before this filing date? Find Prior Art

Description

Artificial intelligence-based diabetic foot disease prediction device and method

[0001] The present invention relates to an artificial intelligence-based device and method for predicting diabetic foot disease. For example, the present invention relates to a technique for utilizing an artificial intelligence model to perform early prediction of diabetic foot lesions in response to a super-aged society.

[0002] The global burden of Type 2 diabetes continues to increase. Particularly in super-aged societies, diabetic peripheral neuropathy (DPN) is the most common complication, occurring in approximately 50% of diabetic patients. However, about half of diabetic patients with DPN experience no symptoms until the disease becomes severe, which acts as a major factor making early detection in the early stages difficult.

[0003] In addition, if diabetic peripheral neuropathy is not detected early, it can progress to diabetic foot lesions, which can ultimately lead to foot ulcers and amputation. Therefore, regular screening for diabetic peripheral neuropathy is recommended in clinical guidelines.

[0004] Currently, traditional diagnostic methods used to diagnose diabetic peripheral neuropathy in clinical settings include the Superficial Pain Test, Monofilament Test, and Tuning Fork Test. However, these traditional methods have several limitations.

[0005] First, these tests are subjective and heavily dependent on patient responses; second, they are not tailored to the patient's demographic characteristics, making personalized diagnosis difficult. Furthermore, they have low sensitivity as they can primarily detect disease in advanced stages, and it is difficult to pinpoint the exact location of peripheral nerve damage. Consequently, traditional diagnostic methods lack reliability and can only detect neuropathy at advanced stages.

[0006] Meanwhile, although Nerve Conduction Studies (NCS) are considered the gold standard for diagnosing diabetic peripheral neuropathy, they are difficult to access in all medical settings due to their high cost and the need for specialized equipment and personnel. Therefore, the development of cost-effective and scalable diagnostic strategies is essential for the early detection and management of diabetic peripheral neuropathy.

[0007] The technology forming the background of this invention is disclosed in Korean Registered Patent Publication No. 10-2535884.

[0008] The present invention aims to solve the problems of the aforementioned conventional technology by providing an artificial intelligence-based method for predicting diabetic foot disease that can predict diabetic peripheral neuropathy early through a multimodal deep learning model combining foot radiographic images and electronic medical record data.

[0009] The present invention aims to solve the problems of the aforementioned conventional technology by providing a predictive model that enables early intervention by assessing the risk of developing diabetic foot disease in a clinically accessible and cost-effective manner.

[0010] However, the technical problems that the embodiments of the present invention aim to solve are not limited to the technical problems described above, and other technical problems may exist.

[0011] As a technical means for achieving the above-mentioned technical problem, an artificial intelligence-based method for predicting diabetic foot disease according to one embodiment of the present invention may include: (a) collecting training data including image data of foot areas and electronic medical record data obtained from a plurality of patients; (b) extracting image features from the image data among the training data and extracting clinical features from the electronic medical record data among the training data to form a multimodal data set combining the image features and the clinical features; (c) training an artificial intelligence-based disease prediction model using the multimodal data set; (d) acquiring target data including image data and electronic medical record data of a subject for analysis; and (e) inputting the target data into the disease prediction model to output prediction information including the probability of onset of diabetic foot disease of the subject for analysis.

[0012] In addition, an artificial intelligence-based method for predicting diabetic foot disease according to one embodiment of the present invention may include, after step (a), a step of augmenting the image data by dividing the image data according to a preset anatomical region that includes the entire foot area, the forefoot, and the hindfoot.

[0013] In addition, the above diabetic foot disease may include diabetic peripheral neuropathy (DPN).

[0014] In addition, the above diabetic foot disease may include diabetic foot lesions.

[0015] In addition, an artificial intelligence-based method for predicting diabetic foot disease according to one embodiment of the present invention may include (f) a step of visualizing and outputting information regarding a local region of the target data that the disease prediction model primarily referenced to derive the prediction information.

[0016] In addition, the above disease prediction model may include a deep learning model that includes a residual neural network structure.

[0017] In addition, the above residual neural network includes a ResNet50 structure and can use pre-trained weights to extract image features of the foot region.

[0018] Additionally, the above step (b) may include a step of performing feature selection based on association with diabetic foot disease among a plurality of clinical features extracted from the electronic medical record data.

[0019] Additionally, the above step (c) may include the step of concatenating the image features and the clinical features to generate an integrated feature vector.

[0020] In addition, the prediction information may include the result of classifying the subject of analysis as having the diabetic foot disease or as normal without the diabetic foot disease, and the probability value for the result.

[0021] Meanwhile, a method for training an artificial intelligence model for predicting diabetic foot disease according to one embodiment of the present invention may include: (a) collecting training data including image data of foot areas obtained from a plurality of patients and electronic medical record data; (b) extracting image features from the image data among the training data and extracting clinical features from the electronic medical record data among the training data to form a multimodal data set combining the image features and the clinical features; and (c) training an artificial intelligence-based disease prediction model using the multimodal data set.

[0022] In addition, the disease prediction model may be constructed to output predictive information including the likelihood of developing diabetic foot disease in the subject of analysis, based on target data including image data and electronic medical record data of the subject of analysis.

[0023] Meanwhile, an artificial intelligence-based diabetic foot disease prediction device according to one embodiment of the present invention may include: a data collection unit that collects training data including image data of foot areas and electronic medical record data obtained from a plurality of patients; a data processing unit that extracts image features from the image data among the training data and extracts clinical features from the electronic medical record data among the training data to form a multimodal data set combining the image features and the clinical features; a model training unit that trains an artificial intelligence-based disease prediction model using the multimodal data set; and an analysis execution unit that acquires target data including image data and electronic medical record data of an analysis subject, inputs the target data into the disease prediction model, and outputs prediction information including the probability of onset of diabetic foot disease of the analysis subject.

[0024] In addition, the data processing unit can augment the image data by dividing the image data according to anatomical regions pre-set to include the entire foot area, the forefoot, and the hindfoot.

[0025] In addition, an artificial intelligence-based diabetic foot disease prediction device according to one embodiment of the present invention may include a result interpretation unit that visualizes and outputs information regarding a local region of the target data that the disease prediction model primarily referenced to derive the prediction information.

[0026] In addition, the data processing unit may perform feature selection based on the association with diabetic foot disease among a plurality of clinical features extracted from the electronic medical record data.

[0027] In addition, the above model learning unit can generate an integrated feature vector by concatenating the above image features and the above clinical features.

[0028] Meanwhile, a learning device for an artificial intelligence model for predicting diabetic foot disease according to one embodiment of the present invention may include a data collection unit that collects learning data including image data of foot areas and electronic medical record data obtained from a plurality of patients; a data processing unit that extracts image features from the image data among the learning data and extracts clinical features from the electronic medical record data among the learning data to form a multimodal data set combining the image features and the clinical features; and a model learning unit that trains an artificial intelligence-based disease prediction model using the multimodal data set.

[0029] The means for solving the problem described above are merely exemplary and should not be interpreted as intended to limit the present invention. In addition to the exemplary embodiments described above, additional embodiments may exist in the drawings and the detailed description of the invention.

[0030] According to the solution to the problem of the present invention described above, an artificial intelligence-based method for predicting diabetic foot disease can be provided, which can predict diabetic peripheral neuropathy early through a multimodal deep learning model that combines foot radiographic images and electronic medical record data.

[0031] According to the aforementioned means for solving the problem of the present invention, it is possible to provide a predictive model that enables early intervention by assessing the risk of developing diabetic foot disease in a clinically accessible and cost-effective manner.

[0032] According to the solution to the problem of the present invention described above, a multimodal deep learning model integrating foot radiographic images and electronic medical record data can not only predict diabetic foot disease with high accuracy but also demonstrate superior generalization performance compared to conventional single-modal approaches while reducing the problem of overfitting.

[0033] According to the solution to the problem of the present invention described above, by utilizing general radiographic images and electronic medical records without the need for separate examination equipment, it can be used as an accessible and cost-effective tool for the early detection of diabetic foot disease even in resource-constrained environments.

[0034] However, the effects obtainable from this invention are not limited to those described above, and other effects may exist.

[0035] FIG. 1 is a schematic diagram of an artificial intelligence-based diabetic foot disease prediction system according to one embodiment of the present invention.

[0036] FIG. 2 is a conceptual diagram illustrating the operation process of an artificial intelligence-based diabetic foot disease prediction system according to one embodiment of the present invention.

[0037] Figure 3 is a diagram illustrating exemplary training data collected to include image data and electronic medical record data.

[0038] Figure 4 is a conceptual diagram illustrating data augmentation for image data.

[0039] Figure 5 is a conceptual diagram showing the architecture of an artificial intelligence-based disease prediction model built using a multimodal dataset.

[0040] Figure 6 is a conceptual diagram showing the detailed structure of the residual neural network of an artificial intelligence-based disease prediction model.

[0041] Figure 7 is a diagram illustrating an example of an interface that visualizes and displays the major local areas referenced when deriving prediction information.

[0042] FIGS. 8a to 8e are experimental examples linked to an artificial intelligence-based diabetic foot disease prediction device according to one embodiment of the present invention, and are drawings showing a comparison of the performance of a disease prediction model.

[0043] FIG. 9 is a schematic diagram of an artificial intelligence-based diabetic foot disease prediction device according to one embodiment of the present invention.

[0044] FIG. 10 is a flowchart of an operation for an artificial intelligence-based diabetic foot disease prediction method according to one embodiment of the present invention.

[0045] Embodiments of the present invention are described below with reference to the attached drawings to enable those skilled in the art to easily implement the invention. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0046] Throughout this specification, when a part is described as being "connected" to another part, this includes not only cases where they are "directly connected," but also cases where they are "electrically connected" or "indirectly connected" with other elements interposed between them.

[0047] Throughout the entire specification, when a component is described as being located "on," "on top," "on top," "under," "on bottom," or "on bottom" of another component, this includes not only cases where the component is in contact with the other component but also cases where another component exists between the two components.

[0048] Throughout this specification, when a part is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0049] The present invention relates to an artificial intelligence-based device and method for predicting diabetic foot disease. For example, the present invention relates to a technique for utilizing an artificial intelligence model to perform early prediction of diabetic foot lesions in response to a super-aged society.

[0050] FIG. 1 is a schematic diagram of an artificial intelligence-based diabetic foot disease prediction system according to one embodiment of the present invention.

[0051] Referring to FIG. 1, an artificial intelligence-based diabetic foot disease prediction system (10) (hereinafter referred to as the 'diabetic foot disease prediction system (10)') according to one embodiment of the present invention may include an artificial intelligence-based diabetic foot disease prediction device (100) (hereinafter referred to as the 'prediction device (100)'), a shooting device (200), a user terminal (300), and a database (400).

[0052] The prediction device (100), the shooting device (200), the user terminal (300), and the database (400) can communicate with each other through a network (20). The network (20) refers to a connection structure that enables information exchange between each node, such as terminals and servers. Examples of such a network (20) include, but are not limited to, a 3GPP (3rd Generation Partnership Project) network, an LTE (Long Term Evolution) network, a 5G network, a WIMAX (World Interoperability for Microwave Access) network, the Internet, a LAN (Local Area Network), a Wireless LAN (Wireless Local Area Network), a WAN (Wide Area Network), a PAN (Personal Area Network), a Wi-Fi network, a Bluetooth network, a satellite broadcasting network, an analog broadcasting network, and a DMB (Digital Multimedia Broadcasting) network.

[0053] The user terminal (300) can be any type of wireless communication device, such as a smartphone, smartpad, tablet PC, PCS (Personal Communication System), GSM (Global System for Mobile communication), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminal.

[0054] Additionally, in the description of the embodiment of the present invention, the imaging device (200) may include various types of medical imaging equipment for acquiring image data (11) of the foot area of ​​the subject of analysis (1).

[0055] For example, the imaging device (200) may include, but is not limited to, an X-ray imaging device, a magnetic resonance imaging (MRI) device, a computed tomography (CT) device, an ultrasound imaging device, etc.

[0056] According to one embodiment of the present invention, the imaging device (200) may mainly include an X-ray imaging device, which may have the advantage of being easily accessible and cost-effective in a general clinical environment.

[0057] For example, image data (11) acquired from a shooting device (200) can be transmitted to a prediction device (100) and a database (400) via a network (20), and can be stored and processed in various digital image formats (e.g., DICOM (Digital Imaging and Communications in Medicine), JPEG, PNG, etc.). Depending on the type of shooting device (200), characteristics such as resolution, contrast, and brightness of the image data (11) may vary, and the prediction device (100) may be configured to process these various types of image data (11).

[0058] Additionally, in the description of the embodiments of the present invention, the database (400) may be a device or server for storing, managing, and providing various data necessary for the operation of the diabetic foot disease prediction system (10).

[0059] For example, the database (400) can store image data (11) obtained from the imaging device (200), electronic medical record data (12) linked from the Hospital Information System (HIS), parameter values, weights, model structure information, etc. used for learning and inference of the disease prediction model (1000).

[0060] Additionally, the database (400) can store intermediate results generated during the model training process, model performance evaluation metrics, verification results, etc. The database (400) can be implemented in various forms, such as a relational database, a NoSQL database, or an object-oriented database, and can be physically located on a single server or configured as a distributed storage system. The database (400) may include functions such as encryption and access control for data security and personal information protection, and may have an optimized structure for the efficient processing of large volumes of medical data.

[0061] FIG. 2 is a conceptual diagram illustrating the operation process of an artificial intelligence-based diabetic foot disease prediction system according to one embodiment of the present invention.

[0062] Referring to FIG. 2, according to one embodiment of the present invention, in a diabetic foot disease prediction system (10), image data (11) of the foot area of ​​an analysis subject (1) can be obtained through a shooting device (200). At the same time, electronic medical record data (12) including clinical information of the analysis subject (1) can be collected from a hospital information system.

[0063] These image data (11) and electronic medical record data (12) can be input into a prediction device (100) and analyzed through an artificial intelligence-based disease prediction model (1000), and the disease prediction model (1000) can predict the likelihood of developing diabetic foot disease in the subject of analysis (1) based on the input data and output it in the form of prediction information.

[0064] The predicted information output in this way can be transmitted to medical staff (2) through a user terminal (300), and medical staff (2) can refer to it to establish a diagnosis and treatment plan for the subject of analysis (1). This entire process can function as an 'AI-assisted DPN diagnosis' system that assists in the process of early diagnosis and management of diabetic peripheral neuropathy (DPN) using artificial intelligence technology.

[0065] In this regard, the prediction information analyzed by the disease prediction model (1000) according to one embodiment of the present invention may include prediction results for various types of diabetic foot diseases. Major target diseases include diabetic peripheral neuropathy (DPN) and diabetic foot lesions.

[0066] Specifically, diabetic peripheral neuropathy is a condition in which peripheral nerves are damaged due to high blood sugar. It occurs in about 50% of diabetic patients and is characterized by being difficult to detect in the early stages when there are no symptoms. Diabetic foot lesions include ulcers, infections, and tissue damage occurring in the foot due to the progression of neuropathy, and are complications that can lead to amputation in severe cases.

[0067] Meanwhile, according to an embodiment of the present invention, the disease prediction model (1000) may operate to predict both diseases or to predict them in stages. For example, after first predicting the likelihood of developing diabetic peripheral neuropathy, the risk of developing foot lesions may be further evaluated for patients in whom neuropathy is detected. As another example, the prediction device (100) may apply a selective prediction method to perform predictions only for specific diseases based on patient characteristics or requests from medical staff. In this way, the disease prediction model (1000) can be utilized as a decision support tool for managing complications of diabetic patients in clinical settings by providing prediction information on various diabetic foot diseases.

[0068] Below, the specific functions and operations of the prediction device (100) will be described in detail.

[0069] First, the prediction device (100) can collect learning data including image data (11) of the foot area obtained from a plurality of patients and electronic medical record data (12).

[0070] Figure 3 is a diagram illustrating exemplary training data collected to include image data and electronic medical record data.

[0071] Referring to FIG. 3, in the description of an embodiment of the present invention, image data (11) may include a medical image of the foot of the subject of analysis (1). In the main embodiment of the present invention, a foot radiograph (X-ray) may be used as image data (11), but is not limited thereto and various medical image modalities such as MRI, CT, and ultrasound images may be applied.

[0072] For example, image data (11) may be acquired directly through a shooting device (200) or collected by retrieving images previously stored in a database (400). The image data (11) may include various anatomical parts such as the entire foot area, forefoot, and hindfoot, and may include images taken from various angles (e.g., side, front, back, etc.).

[0073] In one embodiment of the present invention, a foot radiograph that can be easily obtained in a general clinical environment is used as the main image data (11), thereby having the advantage of being able to predict diabetic foot disease early without additional examination costs. The image data (11) can be stored in DICOM format or converted into standard image formats such as JPEG or PNG and processed, and can be used as input to a disease prediction model (1000) after undergoing preprocessing processes such as resolution, brightness, and contrast.

[0074] Additionally, with reference to FIG. 3, in the description of the embodiment of the present invention, the electronic medical record data (12) may include clinical information, demographic information, biomarkers, test results, disease and treatment history, etc. of the subject of analysis (1). Specifically, the electronic medical record data (12) may include age, gender, body mass index (BMI), duration of diabetes, blood sugar levels (e.g., fasting blood sugar, glycated hemoglobin (HbA1c), etc.), cholesterol levels, kidney function indicators, smoking status, drinking status, presence of hypertension, history of cardiovascular disease, pre-existing neuropathy symptoms, etc.

[0075] These electronic medical record data (12) may be extracted directly from an Electronic Medical Record (EMR) system within a hospital information system or collected through a hospital terminal (not shown) linked to a database (400). The electronic medical record data (12) may follow medical data standard formats such as HL7 (Health Level Seven) and FHIR (Fast Healthcare Interoperability Resources), and processing such as anonymization and encryption may be applied for data privacy and security.

[0076] In addition, the electronic medical record data (12) may also reflect information extracted from unstructured medical text (e.g., doctor's opinion, patient symptom record, etc.) through Natural Language Processing (NLP) technology.

[0077] Additionally, referring to FIG. 3, in an embodiment of the present invention, the training data may consist of 607 foot radiographs and 133 EMR data obtained from 133 patients, but is not limited thereto.

[0078] In this training data set, image data (11) and electronic medical record data (12) are interconnected through patient identifiers. Additionally, the training data is labeled into "DPN(+)" (diabetic peripheral neuropathy positive) and "DPN(-)" (diabetic peripheral neuropathy negative) groups based on the presence or absence of diabetic peripheral neuropathy. These labels are assigned based on objective diagnostic criteria, such as Nerve Conduction Studies (NCS) results, and are essential elements for the disease prediction model (1000) to be trained using a supervised learning method.

[0079] Figure 4 is a conceptual diagram illustrating data augmentation for image data.

[0080] Referring to FIG. 4, the prediction device (100) can augment the image data (11) collected as training data by dividing the image data (11) according to anatomical regions pre-set to include the entire foot area, the forefoot, and the hindfoot.

[0081] Specifically, in this regard, the left side of FIG. 4 shows a symptom site of the foot area, FIG. 4 (a) is image data (11) corresponding to the original image, FIG. 4 (b) shows a processed image cropped from the original image to include the entire foot area, FIG. 4 (c) shows a processed image cropped from the original image to include the forefoot, and FIG. 4 (d) shows a processed image cropped from the original image to include the hindfoot.

[0082] In this regard, the prediction device (100) disclosed herein can improve learning performance from limited image data (11) through a data augmentation method. Specifically, the original image is divided into three anatomical regions: the entire foot (Foot), the forefoot, and the hindfoot, and each is generated as an independent processed image. During this division process, the augmentation ratio can be varied according to the clinical importance of each anatomical region.

[0083] For example, the forefoot region, where diabetic peripheral neuropathy is mainly observed, can be augmented with a larger number of images. Specifically, differential augmentation is possible by generating one processed image of the entire foot, two images of the forefoot region, and one image of the hindfoot region per original image.

[0084] As another example, the prediction device (100) can maximize the data augmentation effect by performing cropping with different sizes and ratios even within the same anatomical region. For example, in the forefoot region, both a narrow crop focused on the toes and a wide crop including the entire metatarsal can be generated, but is not limited to this. Such data augmentation techniques can be understood as ensuring diversity in the model's training data and improving the ability to recognize specific anatomical structures, thereby ultimately enhancing prediction performance.

[0085] Additionally, the prediction device (100) can extract image features from image data (11) among the collected training data and extract clinical features from electronic medical record data (12) among the collected training data to form a multimodal data set that combines the extracted image features and clinical features.

[0086] In this regard, in the description of the embodiments of the present invention, a 'multimodal dataset' refers to a dataset constructed by integrating different types of data (modalities), namely imaging features and clinical features. This multimodal approach has the advantage of providing more comprehensive and diverse information to the model compared to a method using only a single type of data.

[0087] Specifically, in the process of constructing a multimodal data set, visual features are first extracted from image data (11) through a Convolutional Neural Network (CNN). At the same time, clinical features are constructed by selecting and preprocessing clinically important variables from electronic medical record data (12). Afterward, the two sets of features are combined in an appropriate manner and integrated into a single input vector or tensor.

[0088] In this integration process, techniques such as normalization and dimension reduction, which consider the scale, dimensions, and characteristics of each modality data, may be applied. The multimodal data set configuration method of the present invention can improve the accuracy and reliability of predicting diabetic foot disease by enabling the simultaneous consideration of information such as the patient's overall health status and disease history, which is difficult to capture with image data alone, and subtle anatomical changes, which are difficult to identify with electronic medical record data alone.

[0089] Figure 5 is a conceptual diagram showing the architecture of an artificial intelligence-based disease prediction model built using a multimodal dataset.

[0090] Referring to FIG. 5, the prediction device (100) can perform feature selection based on association with diabetic foot disease among a plurality of clinical features extracted from electronic medical record data (12).

[0091] Specifically, a feature selection process can be performed to select features that are highly associated with diabetic foot disease among the clinical features extracted from electronic medical record data (12). This feature selection process can be carried out using statistical methods such as forward selection, backward elimination, and stepwise selection, or machine learning-based importance evaluation methods such as Random Forest and Lasso Regression.

[0092] Meanwhile, in an embodiment of the present invention, the duration of diabetes, glycated hemoglobin (HbA1c) levels, body mass index (BMI), triglyceride levels, total cholesterol levels, the presence of reduced foot sensation, and the presence of chronic kidney disease (CKD) can be identified and selected as features that have a high association with diabetic peripheral neuropathy.

[0093] Furthermore, during the feature selection process, not only the individual importance of each feature but also correlations and multicollinearity between features are considered to determine the optimal feature combination. Additionally, missing values ​​are handled through imputation using the mean, median, or multiple imputation methods. This feature selection process reduces model complexity, prevents overfitting, increases computational efficiency, and, above all, contributes to building a clinically interpretable predictive model.

[0094] Additionally, the prediction device (100) can train an artificial intelligence-based disease prediction model (1000) using a multimodal data set.

[0095] Specifically, the prediction device (100) can train a disease prediction model (1000) including a deep learning model having a residual neural network (101) structure.

[0096] Additionally, referring to FIG. 5, the prediction device (100) may construct a disease prediction model (1000) that includes a ResNet50 structure and a residual neural network (101) that uses pre-trained weights to extract image features of the foot region.

[0097] Additionally, referring to FIG. 5, the prediction device (100) can generate an integrated feature vector by concatenating image features and clinical features extracted from image data (11) and electronic medical record data (12), respectively.

[0098] Meanwhile, according to one embodiment of the present invention, the prediction device (100) may apply a dynamic weighting adjustment mechanism during the integration process of image features and clinical features, and the mechanism may operate in a manner that automatically adjusts the relative importance of image data (11) and electronic medical record data (12) according to the characteristics of the subject of analysis (1).

[0099] Specifically, the prediction device (100) can learn and adjust weights between two modality data by utilizing a self-attention module. For example, if the foot image shows clear abnormal findings, a higher weight may be assigned to the image features, and if there are only subtle changes in the image but high-risk factors (e.g., rapid increase in blood sugar, complaints of sensory abnormalities, etc.) are found in the electronic medical record, a higher weight may be assigned to the clinical features.

[0100] Furthermore, according to one embodiment of the present invention, the aforementioned dynamic weight adjustment can be implemented by utilizing the Fisher Information Matrix to calculate the uncertainty of each modality feature and assigning higher weights to modalities with lower uncertainty (in other words, relatively more reliable). Additionally, weight adjustment in the temporal dimension is also possible. In the case where there is recently captured image data and clinical data recorded a long time ago, higher weights can be assigned to the recent data by considering temporal proximity. This dynamic weight adjustment mechanism improves the prediction accuracy of the model and enables analysis more customized to the individual patient's situation.

[0101] Figure 6 is a conceptual diagram showing the detailed structure of the residual neural network of an artificial intelligence-based disease prediction model.

[0102] Referring to FIG. 6, the residual neural network (101) may include a ResNet50 structure. ResNet50 is a deep convolutional neural network composed of multiple layers, and is an architecture that introduces the concept of residual learning to solve the vanishing gradient problem that occurs in existing deep neural networks. Residual learning is a method of learning a mapping of the form H(x) = F(x) + x by adding skip connections between layers to directly add the input to the output. This structure has the advantage of stabilizing the learning of very deep neural networks and improving performance.

[0103] Specifically, ResNet50 can be composed of convolutional layers, batch normalization layers, activation functions, pooling layers, etc., which are arranged in a repetitive block form. In the embodiment of the present invention, a transfer learning approach can be applied in which weights of a ResNet50 model pre-trained on the ImageNet dataset are used as initial values ​​and fine-tuned to fit foot image data. At this time, various strategies can be used, such as freezing the lower layers and training only the upper layers, or training all layers with various learning rates.

[0104] Meanwhile, the disease prediction model (1000) disclosed herein is not limited to the aforementioned ResNet50, and depending on the implementation example of the present invention, it may broadly include various convolutional neural network architectures such as VGG, Inception, DenseNet, and EfficientNet, or various artificial intelligence-based models (vision models) that are already known or may be developed in the future, such as ViT (Vision Transformer).

[0105] Additionally, the prediction device (100) may acquire target data including image data (11) and electronic medical record data (12) of the subject of analysis (1). Meanwhile, in the description of the embodiments of the present invention, 'target data' refers to a data set input to perform actual prediction using a learned disease prediction model (1000). The target data has the same structure as the training data, but may not have a label regarding the occurrence of diabetic foot disease (in the case of a prediction target) or may include a label for performance evaluation (in the case of test data).

[0106] The target data includes image data (11) of the foot area of ​​the subject of analysis (1) and electronic medical record data (12), which undergo the same preprocessing process as the training data and are converted into a form suitable for model input. In an actual clinical setting, the target data is extracted from the latest test results and medical records of new or existing patients, and the prediction results can be used as reference information to support the diagnosis and treatment decisions of medical staff (2).

[0107] Meanwhile, the prediction device (100) can acquire target data through various channels. For example, when an analysis subject (1) visits a medical institution and takes a foot image through a shooting device (200), the corresponding image data (11) can be transmitted to the prediction device (100) in real time through a network (20).

[0108] As another example, the prediction device (100) can have medical staff (2) directly input electronic medical record data (12) of the subject of analysis (1) or select existing records and transmit them to the prediction device (100) through a terminal connected to the hospital information system.

[0109] As another example, the prediction device (100) can retrieve past image data (11) and electronic medical record data (12) that are stored in advance in the database (400) and use them as target data. This can be useful for monitoring the progress of a patient or for retrospective analysis of existing data.

[0110] Additionally, the prediction device (100) can automatically perform preprocessing steps such as data quality verification, format conversion, and normalization when acquiring target data. For image data (11), resolution adjustment, brightness and contrast normalization, and noise removal may be applied, and for electronic medical record data (12), missing value processing, outlier detection, and unit conversion may be performed. These preprocessing steps can increase prediction accuracy by ensuring that the target data has the same distribution as the training data.

[0111] Additionally, the prediction device (100) can input the acquired target data into a disease prediction model (1000) and output prediction information including the likelihood of developing diabetic foot disease in the subject of analysis (1).

[0112] Specifically, the prediction device (100) can output prediction information including the result of classifying the analysis subject (1) as having diabetic foot disease or as normal without diabetic foot disease, and the probability value for the result, using a disease prediction model (1000).

[0113] In this regard, in one embodiment of the present invention, the prediction information output by the disease prediction model (1000) may be provided in various forms and formats. Basically, the prediction information includes a classification result regarding whether the subject of analysis (1) belongs to a category of 'positive' (diseased person) having diabetic foot disease or 'negative' (normal person) not having disease.

[0114] Along with this, a probability value for the classification result is provided so that the reliability of the classification can be quantitatively evaluated. For example, it may be output in the form of "Positive for diabetic peripheral neuropathy (87.5%)". In addition, the prediction information may include a risk score that expresses the disease risk of the subject of analysis (1) as a continuous score between 0 and 1, which may be visually represented as a gauge or color gradation.

[0115] Furthermore, the information may be provided by classifying it into various risk grades (e.g., low, medium, high), or may include information comparing relative risk with patient groups of the same age or with similar clinical characteristics. Additionally, the predictive information may include prediction results in the form of a survival curve indicating changes in the likelihood of disease occurrence over time, which may be particularly useful for establishing long-term patient management plans. These various forms of predictive information can be provided to medical staff (2) in an intuitive and easy-to-understand manner through the interface of the user terminal (300).

[0116] In addition, according to one embodiment of the present invention, the prediction device (100) can implement a personalized model selection mechanism for selecting an optimal disease prediction model (1000) based on the personal characteristics, disease history, biomarkers, etc. of the subject of analysis (1). Specifically, the prediction device (100) can pre-build and store multiple models optimized for different patient groups.

[0117] For example, various models based on characteristics can be constructed, such as a model optimized for the elderly (65 years or older) group, a model optimized for the middle-aged (40-64 years) group, a model optimized for the patient group with a diabetes history of 10 years or more, and a model optimized for the patient group with comorbid cardiovascular disease. When electronic medical record data (12) of the subject of analysis (1) is input, the prediction device (100) automatically selects the model most suitable for the characteristics of the subject through a decision tree-based model selection algorithm.

[0118] At this time, the criteria for model selection may include various clinical variables such as age, gender, body mass index (BMI), duration of diabetes, glycated hemoglobin (HbA1c) levels, and the presence or absence of comorbidities. For example, in the case of an analysis subject (1) who is female, 65 years of age or older, and has had diabetes for 15 years or more, the 'elderly female long-term diabetic patient' model may be automatically selected. In addition, the weights of each characteristic model are adjusted to respond more sensitively to specific patterns or risk factors appearing in the corresponding patient group, thereby providing more accurate prediction results. This personalized model selection mechanism enables more effective response to various patient characteristics compared to a general single-model approach.

[0119] Figure 7 is a diagram illustrating an example of an interface that visualizes and displays the major local areas referenced when deriving prediction information.

[0120] Referring to FIG. 7, the prediction device (100) can visualize and output information about a local area of ​​target data that the disease prediction model (1000) primarily referenced to derive prediction information.

[0121] Specifically, the prediction device (100) disclosed herein can visualize the decision-making process of a model by utilizing the Gradient-weighted Class Activation Mapping (Grad-CAM) technique. Grad-CAM is a technique that calculates the importance of each feature map in the last convolutional layer of a convolutional neural network through class-specific gradients, and based on this, visualizes which area of ​​the input image has a significant influence on the prediction in the form of a heat map.

[0122] In an embodiment of the present invention, Grad-CAM can be used to indicate which part of the foot image the disease prediction model (1000) focused on when predicting diabetic foot disease. This visualization is primarily expressed as a color spectrum ranging from red to blue, with redder colors indicating an area that had a greater influence on the model's decision-making. By displaying the original image and the heatmap superimposed, medical staff (2) can easily identify the anatomical structures focused on by the model.

[0123] In addition, in one embodiment of the present invention, various explainable AI (XAI) techniques may be applied in addition to the aforementioned Grad-CAM. For example, the contribution of each feature may be analyzed through methods such as LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations), or an attention map may be used to visualize which part of the image the model paid attention to. These explanatory functions can increase the reliability of the model's prediction results and assist medical staff (2) in appropriately interpreting the AI ​​model's judgment in a clinical context.

[0124] Meanwhile, according to one embodiment of the present invention, a multi-level visualization technique may be applied to assist in the interpretation of prediction results. This approach visualizes the decision-making process of the model through multiple levels of detail, ranging from overall foot images to detailed anatomical structures.

[0125] For example, in the first level, Grad-CAM displays a heatmap of the key areas across the entire foot that the model has focused on. In the second level, the Region of Interest (ROI) is automatically detected and magnified, and a more detailed heatmap is generated for that area. In the third level, specific anatomical structures (e.g., metatarsals, phalanges, joints, etc.) within the ROI are automatically segmented and labeled, enabling clear identification of which specific structures the model responded to.

[0126] At this time, a separate instance segmentation model, such as U-Net or Mask R-CNN, may be used to segment the anatomical structure of the foot. Additionally, multi-level visualization may include a temporal dimension. If previous images of the subject (1) are available, the current image and the previous image can be compared to highlight the changed parts and visualize how the model responded to these changes. This is particularly useful for monitoring the progression of the disease. This multi-level visualization approach mitigates the 'black box' nature of the model and supports medical staff (2) in understanding the AI ​​model-based prediction results more intuitively and effectively integrating them into clinical decision-making.

[0127] In addition, according to one embodiment of the present invention, the prediction device (100) can provide not only the diagnosis of diabetic foot disease but also the prognosis prediction and complication risk assessment functions. Specifically, the disease prediction model (1000) may have a multi-output structure that predicts the risk of specific complications that may occur within a predetermined monitoring period (e.g., 6 months, 1 year, 3 years, etc.) in addition to basic disease classification.

[0128] For example, complications to be predicted may include the occurrence of foot ulcers, progression of gangrene, the need for amputation, and the risk of infection. Such prognosis prediction can be implemented by applying time series modeling techniques, and disease progression patterns over time can be learned through Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) networks.

[0129] Additionally, the prediction device (100) can map potential pain sites and pressure distributions based on visualization results. To do this, the region of interest identified by the model is overlaid with an anatomical foot pressure map to predict the pain sites the patient may experience or the locations where additional pressure damage may occur, and this information can be used for designing a customized orthotic device or prescribing a special shoe.

[0130] As another example, the prediction device (100) can assist medical staff (2) in establishing an optimal treatment strategy for each patient by providing a function to simulate the potential impact of various therapeutic interventions (e.g., improvement of blood sugar control, prescription of specific drugs, changes in lifestyle, etc.) applicable to the subject of analysis (1) on the predicted risk based on the derived prediction information.

[0131] Hereinafter, with reference to FIGS. 8a to 8e, the results of the performance analysis of the disease prediction model (1000) will be explained as an experimental example (simulation example) linked with the prediction device (100).

[0132] FIGS. 8a to 8e are experimental examples linked to an artificial intelligence-based diabetic foot disease prediction device according to one embodiment of the present invention, and are drawings showing a comparison of the performance of a disease prediction model. Specifically, FIGS. 8a to 8c are conceptual diagrams for comparing the performance of a disease prediction model (1000) according to a learning strategy, and FIGS. 8d and 8e are graphs showing a comparison of the performance of a disease prediction model (1000) according to a model type.

[0133] Specifically, Figure 8a shows the results labeled 'Scratch', which illustrates the learning curve when the model is trained from scratch without prior training. In this case, the loss gradually decreases and the accuracy increases on the training dataset, but on the validation dataset, an overfitting phenomenon is observed where performance improvement is limited or even deteriorates after a certain point.

[0134] Figure 8b shows the result labeled 'Pre-trained(Freeze)', which represents the case where the lower layer weights of a pre-trained ResNet50 model are frozen and only the upper layer is trained. This approach shows fast convergence in the early stages of training, but the final performance may be limited because the model does not sufficiently learn the specific features of the foot images.

[0135] Figure 8c shows the result labeled 'Fine-tuning,' in which pre-trained weights are used as initial values, but the entire layer is fine-tuned using a differential learning rate. This method demonstrates the best learning efficiency and generalization performance among the three approaches. The 'Fine-tuning' strategy enables effective recognition of specific patterns in foot images while maintaining general visual feature extraction capabilities by training the lower layer with a small learning rate and the upper layer with a relatively large learning rate.

[0136] Additionally, referring to FIGS. 8d and 8e, the graphs compare various performance indicators such as AUC (Area Under the ROC Curve), Accuracy, Precision, and Recall according to model configuration. The model configurations are classified into three types: 'Image-only' (using only image data), 'Clinical-only' (using only clinical data), and 'Multimodal' (multimodal approach). The 'Image-only' model uses only foot image data (11) as input, and the 'Clinical-only' model uses only clinical features extracted from electronic medical record data (12). The 'Multimodal' model is a major embodiment of the present invention that utilizes both data types.

[0137] The graph confirms that the 'Multimodal' model demonstrates superior results compared to unimodal models across most performance metrics. Of particular note is that the 'Multimodal' model maintains relatively high performance. This suggests that the multimodal approach possesses more robust generalization capabilities.

[0138] In summary, the multimodal ResNet50 model demonstrated excellent performance with an AUC of 0.894 (95% confidence interval: 0.858-0.930) and a recall of 0.814 (0.751-0.869). Additionally, it is a significant finding that the multimodal model exhibits balanced performance across all performance metrics (AUC, accuracy, precision, recall, F1 score, etc.). This distinguishes it from single-modal models that show superior performance in only specific metrics.

[0139] From a clinical perspective, this balanced performance enhances its value as a practical diagnostic tool that minimizes unnecessary additional tests (adequate precision) while ensuring no positive cases are missed (high recall). In conclusion, the experimental results demonstrate that the multimodal approach proposed by our institution, which integrates foot radiography and electronic medical record data, is effective for the early prediction of diabetic foot disease and, in particular, exhibits excellent generalization performance in real-world clinical settings.

[0140] FIG. 9 is a schematic diagram of an artificial intelligence-based diabetic foot disease prediction device according to one embodiment of the present invention.

[0141] Referring to FIG. 9, the prediction device (100) may include a data collection unit (110), a data processing unit (120), a model learning unit (130), an analysis execution unit (140), and a result interpretation unit (150).

[0142] The data collection unit (110) can collect learning data including image data (11) of the foot area obtained from multiple patients and electronic medical record data (12).

[0143] The data processing unit (120) can augment the image data (11) collected as training data by dividing the image data (11) according to anatomical regions pre-set to include the entire foot area, the forefoot, and the hindfoot.

[0144] Additionally, the data processing unit (120) can extract image features from image data (11) among the collected training data and extract clinical features from electronic medical record data (12) among the collected training data to form a multimodal data set that combines the extracted image features and clinical features.

[0145] Specifically, the data processing unit (120) can perform feature selection based on association with diabetic foot disease among a plurality of clinical features extracted from electronic medical record data (12).

[0146] The model learning unit (130) can train an artificial intelligence-based disease prediction model (1000) using a multimodal data set.

[0147] Specifically, the model learning unit (130) can train a disease prediction model (1000) including a deep learning model having a residual neural network (101) structure.

[0148] In this regard, according to one embodiment of the present invention, the model learning unit (130) may construct a disease prediction model (1000) that includes a ResNet50 structure and a residual neural network (101) that uses pre-trained weights to extract image features of the foot region.

[0149] Additionally, the model learning unit (130) can generate an integrated feature vector by concatenating image features and clinical features extracted from image data (11) and electronic medical record data (12), respectively.

[0150] The analysis execution unit (140) can acquire target data including image data (11) and electronic medical record data (12) of the subject of analysis (1). Additionally, the analysis execution unit (140) can input the acquired target data into a disease prediction model (1000) to output prediction information including the likelihood of developing diabetic foot disease in the subject of analysis (1).

[0151] Specifically, the analysis execution unit (140) can output prediction information including the result of classifying the analysis subject (1) as having diabetic foot disease or as normal without diabetic foot disease, and the probability value for the result, using a disease prediction model (1000).

[0152] The result interpretation unit (150) can visualize and output information about the local region of the target data that the disease prediction model (1000) primarily referenced to derive prediction information.

[0153] Below, based on the details described above, we will briefly examine the operation flow of the present invention.

[0154] FIG. 10 is a flowchart of an operation for an artificial intelligence-based diabetic foot disease prediction method according to one embodiment of the present invention.

[0155] The artificial intelligence-based diabetic foot disease prediction method illustrated in FIG. 10 can be performed by the prediction device (100) described above. Therefore, even if the content is omitted below, the description of the prediction device (100) can be equally applied to the description of the artificial intelligence-based diabetic foot disease prediction method.

[0156] Referring to FIG. 10, in step S11, the data collection unit (110) can collect learning data including image data (11) of the foot area obtained from a plurality of patients and electronic medical record data (12).

[0157] Next, in step S12, the data processing unit (120) can augment the image data (11) collected as training data by dividing the image data (11) according to anatomical regions pre-set to include the entire foot area, the forefoot, and the hindfoot.

[0158] Next, in step S13, the data processing unit (120) can extract image features from image data (11) among the collected training data and extract clinical features from electronic medical record data (12) among the collected training data to form a multimodal data set that combines the extracted image features and clinical features.

[0159] Specifically, in step S13, the data processing unit (120) can perform feature selection based on association with diabetic foot disease among a plurality of clinical features extracted from electronic medical record data (12).

[0160] Next, in step S14, the model training unit (130) can train an artificial intelligence-based disease prediction model (1000) using a multimodal data set.

[0161] Specifically, in step S14, the model learning unit (130) can train a disease prediction model (1000) including a deep learning model having a residual neural network (101) structure.

[0162] In this regard, according to one embodiment of the present invention, in step S14, the model learning unit (130) may construct a disease prediction model (1000) that includes a ResNet50 structure and a residual neural network (101) that uses pre-trained weights to extract image features of the foot region.

[0163] Additionally, in step S14, the model learning unit (130) can generate an integrated feature vector by concatenating image features and clinical features extracted from image data (11) and electronic medical record data (12), respectively.

[0164] Next, in step S15, the analysis execution unit (140) can obtain target data including image data (11) and electronic medical record data (12) of the subject of analysis (1).

[0165] Next, in step S16, the analysis execution unit (140) inputs the acquired target data into a disease prediction model (1000) and outputs prediction information including the likelihood of developing diabetic foot disease in the subject of analysis (1).

[0166] Specifically, in step S16, the analysis execution unit (140) can output prediction information including the result of classifying the analysis subject (1) as having diabetic foot disease or as normal without diabetic foot disease, and the probability value for the result, using a disease prediction model (1000).

[0167] Next, in step S17, the result interpretation unit (150) can visualize and output information about the local region of the target data that the disease prediction model (1000) primarily referenced to derive prediction information.

[0168] In the description above, steps S11 through S170 may be further divided into additional steps or combined into fewer steps according to an embodiment of the present invention. Additionally, some steps may be omitted as necessary, and the order between steps may be changed.

[0169] An artificial intelligence-based method for predicting diabetic foot disease according to one embodiment of the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either individually or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the present invention, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

[0170] In addition, the aforementioned artificial intelligence-based method for predicting diabetic foot disease can also be implemented in the form of a computer program or application executed by a computer stored on a recording medium.

[0171] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical concept or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0172] The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and the concept of equivalents thereof should be interpreted as being included within the scope of the present invention.

[0173] [Explanation of the symbol]

[0174] 10: AI-based diabetic foot disease prediction system

[0175] 100: AI-based diabetic foot disease prediction device

[0176] 110: Data Collection Unit

[0177] 120: Data Processing Department

[0178] 130: Model learning section

[0179] 140: Analysis Execution Unit

[0180] 150: Result Interpretation Section

[0181] 200: Shooting device

[0182] 300: User terminal

[0183] 400: Database

[0184] 20: Network

[0185] 1: Subject of analysis

[0186] 2: Medical staff

[0187] 11: Image data

[0188] 12: Electronic Medical Record Data

[0189] 1000: Disease Prediction Model

[0190] 101: Residual Neural Network

Claims

1. In an artificial intelligence-based method for predicting diabetic foot disease, (a) A step of collecting training data including image data of the foot region and electronic medical record data obtained from multiple patients; (b) a step of extracting image features from the image data among the training data and extracting clinical features from the electronic medical record data among the training data to form a multimodal data set that combines the image features and the clinical features; (c) a step of training an artificial intelligence-based disease prediction model using the above multimodal data set; (d) a step of acquiring target data including image data and electronic medical record data of the subject of analysis; and (e) inputting the target data into the disease prediction model to output prediction information including the likelihood of developing diabetic foot disease in the subject of analysis, A prediction method including 2. In Paragraph 1, After step (a) above, A step of augmenting the image data by dividing the image data according to a preset anatomical region that includes the entire foot area, the forefoot, and the hindfoot, A prediction method that further includes 3. In Paragraph 1, A prediction method characterized in that the above-mentioned diabetic foot disease includes diabetic peripheral neuropathy (DPN).

4. In Paragraph 1, A prediction method characterized in that the above-mentioned diabetic foot disease includes diabetic foot lesions.

5. In Paragraph 1, (f) A step of visualizing and outputting information on a local region of the target data that the disease prediction model primarily referenced to derive the prediction information, A prediction method that further includes 6. In Paragraph 1, The above disease prediction model is, A prediction method characterized by including a deep learning model that includes a residual neural network structure.

7. In Paragraph 6, A prediction method characterized in that the above residual neural network includes a ResNet50 structure and uses pre-trained weights to extract image features of the foot region.

8. In Paragraph 1, The above step (b) is, A step of performing feature selection based on association with diabetic foot disease among a plurality of clinical features extracted from the electronic medical record data, A prediction method that includes 9. In Paragraph 1, The above step (c) is, A step of generating an integrated feature vector by concatenating the above image features and the above clinical features, A prediction method that includes 10. In Paragraph 1, The above prediction information is, A prediction method characterized by including a result of classifying the above-mentioned analysis subject as having the above-mentioned diabetic foot disease or as normal without the above-mentioned diabetic foot disease, and a probability value for the above result.

11. In the training method of an artificial intelligence model for predicting diabetic foot disease, (a) A step of collecting training data including image data of the foot region and electronic medical record data obtained from multiple patients; (b) a step of extracting image features from the image data among the training data and extracting clinical features from the electronic medical record data among the training data to form a multimodal data set that combines the image features and the clinical features; and (c) A step of training an artificial intelligence-based disease prediction model using the above multimodal data set, Includes, The above disease prediction model is, A learning method constructed to output predictive information including the likelihood of developing diabetic foot disease in a subject based on target data including image data and electronic medical record data of the subject being analyzed.

12. In an artificial intelligence-based diabetic foot disease prediction device, A data collection unit that collects learning data including image data of the foot region and electronic medical record data obtained from multiple patients; A data processing unit that extracts image features from the image data among the above training data and extracts clinical features from the electronic medical record data among the above training data to form a multimodal data set combining the image features and the clinical features; A model training unit that trains an artificial intelligence-based disease prediction model using the above multimodal data set; and An analysis execution unit that acquires target data including image data and electronic medical record data of an analysis subject, inputs the target data into a disease prediction model, and outputs prediction information including the likelihood of developing diabetic foot disease of the analysis subject. A prediction device including 13. In Paragraph 12, The above data processing unit is, A prediction device that augments the image data by dividing the image data according to anatomical regions pre-set to include the entire foot area, the forefoot, and the hindfoot.

14. In Paragraph 12, A prediction device wherein the above-mentioned diabetic foot disease includes at least one of diabetic peripheral neuropathy (DPN) and diabetic foot lesions.

15. In Paragraph 12, A result interpretation unit that visualizes and outputs information regarding a local region of the target data that the disease prediction model primarily referenced to derive the prediction information, A prediction device that further includes 16. In Paragraph 12, The above data processing unit is, Feature selection is performed based on the association with diabetic foot disease among the plurality of clinical features extracted from the above electronic medical record data, and The above model learning unit is, A prediction device that generates an integrated feature vector by concatenating the above image features and the above clinical features.

17. In Paragraph 12, The above prediction information is, A prediction device characterized by including a result of classifying the above-mentioned analysis subject as having the above-mentioned diabetic foot disease or as normal without the above-mentioned diabetic foot disease, and a probability value for said result.

18. In a learning device for an artificial intelligence model for predicting diabetic foot disease, A data collection unit that collects learning data including image data of the foot region and electronic medical record data obtained from multiple patients; A data processing unit that extracts image features from the image data among the training data and extracts clinical features from the electronic medical record data among the training data to form a multimodal data set combining the image features and the clinical features; and A model training unit that trains an artificial intelligence-based disease prediction model using the above multimodal data set, Includes, The above disease prediction model is, A learning device constructed to output predictive information including the likelihood of developing diabetic foot disease in a subject based on target data including image data and electronic medical record data of the subject.