Prognostic prediction model for immunoglobulin a nephropathy disease

Abstract

The present invention relates to an artificial intelligence model and an implementation method therefor, which are capable of selecting and extracting, from CT images of patients with IgA nephropathy, image features significant for prognostic prediction and applying same to a machine learning model to thereby predict, with high accuracy, the likelihood of the patients with IgA nephropathy progressing to end-stage renal failure within five years. The present invention provides a prognostic prediction model that rapidly predicts the prognosis of a patient without an invasive kidney biopsy, overcomes the limitations of conventional pathology diagnosis relying on invasive methods, and enables periodic prognostic evaluation with significantly improved reliability. In addition, the present invention is capable of precisely reflecting characteristics of each item and accurately predicting a clinical course of a patient, by combining various feature selection methods and binary classifiers for each item of mesangial hypercellularity (M), endothelial hypercellularity (E), segmental glomerulosclerosis (S), and tubular atrophy / interstitial fibrosis (T), which constitute a MEST score.

Description

Prognostic prediction model for immunoglobulin A nephropathy

[0001] The present invention relates to an artificial intelligence model that extracts radioomics features from CT (Computed Tomography) images of a patient with immunoglobulin A nephropathy and predicts the prognosis of said patient based thereon, and to the use thereof.

[0002]

[0003] IgA nephropathy (IgAN) is a type of chronic glomerulonephritis caused by the abnormal deposition of immunoglobulin A in the glomeruli, and is known as the most common primary glomerular disease worldwide. IgA nephropathy primarily affects young adults and initially presents with relatively nonspecific symptoms such as hematuria and proteinuria; however, renal function gradually declines over the long term, eventually progressing to end-stage renal disease (ESRD). In fact, IgA nephropathy is reported to be one of the leading causes of chronic renal failure globally, and it is known that approximately 20–40% of patients reach end-stage renal disease within 20 years of onset. Therefore, accurately predicting the prognosis of patients with IgA nephropathy is essential for establishing treatment strategies and providing personalized patient management.

[0004] Currently, the prognosis assessment of patients with IgA nephropathy relies primarily on the pathological indicator known as the MEST score (Mesangial hypercellularity, Endocapillary hypercellularity, Segmental glomerulosclerosis, Tubular atrophy / interstitial fibrosis). This score is calculated based on histological information obtained through kidney biopsy and is widely used to evaluate disease activity and prognosis in patients with IgA nephropathy. Furthermore, the MEST score serves as an important criterion for predicting the decline in renal function and the risk of progression to end-stage renal failure, and is utilized in various ways in clinical practice.

[0005] However, kidney biopsy is an invasive procedure that places a significant burden on patients, and it has limitations in that repeated examinations are difficult due to the possibility of complications. In particular, since the course of the disease can change even after steroid or immunosuppressant treatment in patients with IgA nephropathy, there is a continuously being raised need for non-invasive and repeatable prognostic prediction technologies that can reflect changes in the condition before and after treatment.

[0006] Meanwhile, CT (computed tomography) imaging is a widely used imaging technique in clinical practice that can quantitatively evaluate tissue structural and functional changes. Recently, there have been active attempts to replace or supplement existing pathological markers by combining radiomics technology, which extracts quantitative features from CT images, with artificial intelligence techniques. However, a technology to predict MEST scores based on CT images of patients with IgA nephropathy and thereby predict the patient's ESRD risk has not yet been established.

[0007] Therefore, there is an urgent need to develop a new method that can replace the MEST score with CT image-based radiomics analysis without invasive biopsy and evaluate patient prognosis by applying it to existing predictive tools.

[0008]

[0009] Conventionally, the prognosis assessment of patients with IgA nephropathy relied on MEST scores obtained from kidney biopsies; however, kidney biopsies are invasive procedures that place a heavy burden on patients and are difficult to repeat. Accordingly, the inventors have made diligent research efforts to implement an artificial intelligence model capable of predicting patient prognosis more quickly and with improved reliability by calculating MEST scores based on CT images of patients with IgA nephropathy. As a result, by extracting and selecting image features significant for calculating MEST scores from CT images of patients with IgA nephropathy and applying them to a machine learning model, a prognosis prediction model was developed that can predict with high accuracy the likelihood of patients with IgA nephropathy progressing to end-stage renal failure within the next five years. By rapidly predicting patient prognosis without invasive kidney biopsies, this invention overcomes the limitations of existing pathological diagnoses and can significantly contribute to establishing personalized treatment strategies for patients with IgA nephropathy.

[0010] Accordingly, the objective of the present invention is to provide a method for providing information necessary for predicting the prognosis of patients with kidney disease.

[0011] Another objective of the present invention is to provide a device for predicting the prognosis of patients with kidney disease.

[0012] Another objective of the present invention is to provide a method for training an artificial intelligence model for predicting the prognosis of patients with kidney disease.

[0013] Another objective of the present invention is to provide a prediction device capable of calculating a MEST score based on a patient's medical images, thereby overcoming the limitations of conventional invasive pathological diagnosis and enabling repeated prognosis prediction for patients with IgA nephropathy.

[0014]

[0015] Other objects and advantages of the present invention will become more apparent from the following detailed description of the invention, claims, and drawings.

[0016]

[0017] According to one aspect of the present invention, the present invention provides a method for providing information necessary for predicting the prognosis of a patient with kidney disease, comprising the following steps:

[0018] (a) A step of extracting digitized image information from a CT image in which the lesion is segmented in a patient with kidney disease;

[0019] (b) a step of performing a feature selection process on the above image information;

[0020] (c) a step of inputting the selected features into a machine learning model to calculate a MEST (Mesangial hypercellularity, Endocapillary hypercellularity, Segmental glomerulosclerosis, Tubular atrophy / Interstitial fibrosis) score; and

[0021] (d) A step of obtaining a prognostic prediction value by inputting the MEST score calculated by the machine learning above into a standard prognostic prediction tool.

[0022] The inventors have made diligent research efforts to discover a method that can more rapidly calculate MEST scores, which have a significant correlation with the prognosis of patients with IgA nephropathy, through a non-invasive method. As a result, they calculated MEST scores that have a positive correlation with poor prognosis from CT images of patients with IgA nephropathy and developed a prognosis prediction model that enables periodic prognosis evaluation without invasive kidney biopsy by utilizing these scores. The artificial intelligence model selects imaging features significant for prognosis prediction from the CT images of patients with nephropathy and predicts with high accuracy the likelihood that the patient will progress to end-stage renal failure within the next five years. Accordingly, the present invention identifies key imaging features that can be utilized for the prognosis evaluation of patients with IgA nephropathy and derives highly reliable results by comparing and evaluating prediction performance using various classification algorithms.

[0023]

[0024] As used in this specification, the term “prognosis” refers to a forecast of future symptoms or course determined by diagnosing a disease. For patients with kidney disease, prognosis generally refers to the rate of decline in renal function, the persistence of proteinuria, the degree of decrease in estimated glomerular filtration rate (eGFR), or progression to end-stage renal failure within the next few years. Prognosis prediction is an important clinical task in establishing patient-specific treatment strategies and developing long-term management plans; in patients with IgA nephropathy, it is known that the pathological indicator MEST score and clinical indicators have a close correlation with prognosis. The term “prognosis” as used in this invention encompasses both positive and negative prognoses. A positive prognosis includes stabilization of renal function, improvement of proteinuria, and maintenance of a stable state of the disease, while a negative prognosis includes decline in renal function, worsening of proteinuria, disease progression, and transition to end-stage renal failure.

[0025]

[0026] As used herein, the term “lesion segmentation” refers to a series of processes for identifying a lesion region of interest in a CT image and defining the spatial boundaries or masks of that region to distinguish the lesion area from other tissues within the entire image and to identify or extract it independently. Lesion segmentation can be performed through an automated artificial intelligence-based image segmentation algorithm, or it may be performed manually based on morphological features, including boundary clarity and differences in tissue density. Furthermore, since it encompasses a series of processes for segmentation based not only on morphological features but also on functional or quantitative image information, including signal intensity and changes over time, “lesion segmentation” in this specification includes all procedures for recognizing a lesion or segmentation region of interest from a CT image and defining that region. Specifically, the present invention can segment a renal lesion region in a CT image using a deep learning segmentation model. More specifically, the present invention can segment a renal lesion region from a CT image using nnU-Net (neural network-based U-Net), a deep learning-based medical image segmentation model. The above automated segmentation method can extract the spatial boundaries of the elongation within the image with high precision, providing high reproducibility and efficiency compared to manual segmentation.

[0027] Accordingly, in this specification, “CT image with segmented lesions” refers to image data in which a lesion or region of interest is identified and segmented from a CT image by a manual, semi-automatic, or automated method, and boundary information or a mask of the lesion is defined, and the image includes a state in which it can be used as base data for quantitative image feature extraction, analysis, or artificial intelligence model training in a subsequent step.

[0028]

[0029] As used in this specification, the term “CT (Computed tomography)” refers to a medical imaging technique that provides high-resolution tomographic images in two or three dimensions by acquiring cross-sectional images of the human body using X-rays and reconstructing them using a computer. CT enables precise observation of the anatomical structures of body organs within an image, and in particular, in the kidney region, it includes various morphological and densitological information that reflects structural changes in the glomeruli, tubules, and interstitium. In the present invention, lesion segmentation is performed using CT images of the kidneys of patients with IgA nephropathy, and quantitative features (radiomics features) within the images are extracted and analyzed based on the resulting image data. These features are then used as input data for an artificial intelligence model to predict the prognosis of the patient.

[0030] As used herein, the term “quantified image information” refers to various quantitative characteristic values ​​extracted from lesions or segmented regions of interest within an image based on medical images, such as CT images. This information includes the image’s intensity, texture, shape, gradient, statistical distribution, and dynamic change, as well as all information that can be extracted from and quantified from CT images. More specifically, quantified image information refers to features reflecting patterns, morphological changes, or functional changes within the image that may correlate with the prognosis of patients with IgA nephropathy. Specifically, said quantified image information may be quantified using open-source image analysis software such as PyRadiomics.

[0031] As used in this specification, the term “feature selection process” refers to a series of procedures for selecting key features that have a significant impact on the prognosis prediction of patients with IgA nephropathy from among quantified image information whose reliability and reproducibility have been confirmed. The feature selection process is performed to significantly improve the accuracy of the prediction model and prevent overfitting, and includes not only all statistical methods including statistical testing and correlation analysis, but also various machine learning-based feature selection methods. Specifically, the feature selection process includes, but is not limited to, CST (Correlation-based Feature Selection), Relief (Relief algorithm), Corr (Correlation-based Selection), LASSO (Least Absolute Shrinkage and Selection Operator), FFS (Forward Feature Selection), RFE (Recursive Feature Elimination), PCA (Principal Component Analysis), and Mutual Information, and may include various feature selection methods commonly used in the industry.

[0032]

[0033] As used in this specification, the term “machine learning” refers to a statistical-based learning algorithm for predicting a patient’s clinical condition based on quantitative information collected from medical images, and includes, but is not limited to, supervised learning, unsupervised learning, and reinforcement learning. Specifically, the machine learning provided by the present invention performs a feature selection process on quantified image information extracted from CT images of patients with IgA nephropathy to calculate a MEST score and predicts the patient’s prognosis with high confidence based on this score.

[0034]

[0035] As used herein, the term “MEST score” refers to the internationally recognized Oxford classification index for evaluating the renal histological characteristics of patients with IgA nephropathy. The MEST score consists of four pathological items, including mesangial hypercellularity, endocapillary hypercellularity, segmental glomerulosclerosis, and tubular atrophy / interstitial fibrosis.

[0036] “M (Mesangial hypercellularity)”, which constitutes the above MEST score, is a phenomenon in which cells abnormally increase in the mesangial region of the glomerulus and is associated with proteinuria and decreased renal function, acting as a signal of worsening prognosis. Specifically, if mesangial hypercellularity is observed in more than 50% of the glomerulus, it is classified as M1, and if not, as M0.

[0037] “E (Endocapillary hypercellularity)”, which constitutes the above MEST score, evaluates the presence of findings of lumen stenosis or occlusion due to proliferation of inflammatory cells and endothelial cells within the glomerular capillaries, and is classified as E1 if one or more such findings are observed in the glomerulus, and E0 if none are observed.

[0038] “S (Segmental glomerulosclerosis)”, which constitutes the above MEST score, evaluates the presence of localized sclerosis or adhesion in a segment of the glomerulus and classifies it as S1 if such findings are observed and S0 if they are not observed.

[0039] “T (Tubular atrophy / Interstitial fibrosis)”, which constitutes the above MEST score, is a quantitative measure of the extent of tubular atrophy and interstitial fibrosis in terms of area. It is classified as T0 when fibrosis has progressed to less than 25% of the renal cortex area, T1 when fibrosis has progressed to 25% or more and 50% or less of the renal cortex area, and T2 when fibrosis has progressed to more than 50%.

[0040] The more unfavorable findings there are in the M, E, S, or T categories, the worse the patient's prognosis tends to be; furthermore, there is a positive correlation between the MEST score and the risk of long-term renal decline and progression to end-stage renal failure in patients with IgA nephropathy. More specifically, the T category is known as the independent prognostic factor most strongly associated with long-term renal decline, while the S category reflects chronic renal damage and disease progression. The M and E categories are associated with proteinuria, active inflammatory response, and treatment responsiveness, and act as important factors determining short-term prognosis in some patients. Therefore, the MEST score serves as a key indicator linking the pathological characteristics and clinical prognosis of patients with IgA nephropathy, playing an essential role in predicting long-term prognosis and establishing treatment strategies.

[0041]

[0042] As used herein, the term “standard prognostic prediction tool” refers to a prediction model commonly used in the diagnostic field to quantitatively evaluate the long-term clinical prognosis of patients with IgA nephropathy. The prediction tool calculates the probability that a patient will experience a decline in renal function or reach end-stage renal failure within the next few years by comprehensively considering clinical information, including the patient’s age, gender, and blood pressure, pathological indicators including the MEST score, glomerular filtration rate, proteinuria levels, and serum creatinine concentration. Specifically, the present invention calculates the probability that a patient with IgA nephropathy will reach end-stage renal failure in the future using the International IgA Nephropathy Prediction Tool (IIgAN-PT). The International IgA Nephropathy Prediction Tool is an algorithm built on multinational data that reliably predicts the patient’s long-term prognosis by considering both clinical and pathological variables.

[0043] The present invention enables precise evaluation of a patient's prognosis without invasive procedures by applying an image-based MEST score extracted from CT images to a standard prognosis prediction tool instead of a pathological MEST score calculated from a kidney biopsy.

[0044]

[0045] According to a specific embodiment of the present invention, the feature selection process of step (b) described above is performed by No FS (NO Feature Selection), Corr (Correlation-based Selection), LASSO (Least Absolute Shrinkage and Selection Operator), or FFS (Forward Feature Selection).

[0046] The term “No FS (NO Feature Selection)” as used in this specification refers to a method of using the entire quantified information as is as an input variable for a machine learning model, and by utilizing all extracted image features for training, feature loss can be prevented.

[0047] As used in this specification, the term “Corr (Correlation-based Selection)” refers to a method of selecting only features that are above a specific threshold or rank highly by evaluating the significance of features based on the correlation coefficient between each feature and the target variable for prediction, and is effective in eliminating duplication among features.

[0048] The term “LASSO (Least Absolute Shrinkage and Selection Operator)” as used in this specification refers to a regression analysis-based feature selection method that automatically selects features by adding an L1 regularization term to the objective function to reduce the regression coefficients of unnecessary features to zero. This method performs variable selection and regularization simultaneously, prevents overfitting in high-dimensional data, and improves the generalization ability of the model.

[0049] As used in this specification, the term “FFS (Forward Feature Selection)” refers to a sequential feature selection technique that determines the optimal feature combination by sequentially adding features that contribute to improving classification performance; it evaluates the performance of a model by repeatedly adding features one by one, even though it initially contains no features.

[0050]

[0051] More specifically, the above feature selection process is performed by the LASSO (Least Absolute Shrinkage and Selection Operator) or FFS (Forward Feature Selection) method.

[0052] As described below, the inventors applied the FFS method or the LASSO method to select imaging features that have a significant correlation with the prognosis of patients with IgA nephropathy, and experimentally identified significantly improved reliability in prognostic prediction performance by combining said imaging features with multiple classifiers. Furthermore, considering the pathological characteristics of each of the four items constituting the MEST score, the inventors applied different combinations of classifier-feature selection methods for each item: mesangial proliferation (M), endothelial cell proliferation (E), segmental glomerular sclerosis (S), and tubular atrophy / interstitial fibrosis (T). Accordingly, the artificial intelligence model provided by the present invention demonstrated excellent results in various performance indicators, including sensitivity, specificity, precision, and AUC (Area Under the Curve), in predicting the prognosis of patients with IgA nephropathy, and can be practically utilized in clinical settings.

[0053]

[0054] According to a specific embodiment of the present invention, the step of inputting the selected features of the above-described step (c) into a machine learning model is performed by a binary classifier selected from the group consisting of Random Forest, Adaptive Boosting, KNN (K-Nearest Neighbors), SGD (Stochastic Gradient Descent), SVM (Support Vector Machine), LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), GNB (Gaussian Naive Bayes), SHAP (SHapley Additive exPlanations), and Decision Tree.

[0055] The term “Adab (Adaptive Boosting)” as used in this specification refers to an ensemble learning method that generates a strong classifier with excellent predictive performance by sequentially combining multiple classifiers through iterative training of weak classifiers, and in each iteration step, the accuracy of the classifier is improved by assigning higher weights to samples that were misclassified in the previous step.

[0056] As used herein, the term “SGD (Stochastic Gradient Descent)” refers to a learning method that iteratively updates weights to minimize the model’s loss function using a randomly selected small data batch (mini-batch) rather than the entire dataset. Due to its high computational efficiency, it can be effectively applied to large-scale datasets or high-dimensional feature spaces.

[0057] The term “LDA (Linear Discriminant Analysis)” as used in this specification refers to a linear classification technique that maximizes the degree of distinction between classes by linearly transforming data in a direction that maximizes between-class variance and minimizes within-class variance.

[0058] The term “RF (Random Forest)” as used in this specification refers to an ensemble-based classification technique that trains multiple decision trees and then performs a final prediction from them through a majority voting method or an average. Each tree is trained based on randomly selected data samples and feature sets, thereby ensuring model diversity and prediction stability.

[0059] As used in this specification, the term “DT (Decision Tree)” refers to a classification technique that performs predictions by representing data in a tree structure through multiple rule-based splits. Each internal node represents a condition regarding a specific feature, and the leaf node ultimately reached by following the branches provides the prediction result.

[0060] The term “SHAP (SHapley Additive exPlanations)” as used in this specification refers to an explanatory technique for interpreting the prediction results of a machine learning model based on Shapley values ​​from game theory. SHAP can transparently explain the internal decision-making of a black-box model by quantitatively calculating the degree to which each input feature contributes to individual prediction values. This allows for an intuitive understanding of how the model determined the prognosis of a specific patient and contributes to identifying clinically important imaging features. More specifically, the step of inputting the selected features of the present invention into a machine learning model is performed by SHAP, DT, or RF.

[0061]

[0062] According to a specific embodiment of the present invention, the aforementioned kidney disease is selected from the group consisting of IgA nephropathy, membranous nephropathy, minimal change nephropathy, diabetic nephropathy, lupus nephritis, polycystic nephropathy, and hypertensive nephropathy.

[0063] As used herein, the term “kidney disease” refers to a condition in which the normal physiological functions of the kidneys, including the excretion of waste products from the body, maintenance of fluid and electrolyte balance, and hormone regulation, are impaired, and includes all diseases resulting from structural or functional abnormalities of the kidneys. Specifically, the kidney disease refers to, but is not limited to, IgA nephropathy, membranous nephropathy, minimal change nephropathy, diabetic nephropathy, lupus nephritis, polycystic nephropathy, and hypertensive nephropathy, and includes all diseases resulting from pathological abnormalities of the glomeruli, tubules, interstitium, or renal vascular system. More specifically, the kidney disease refers to IgA nephropathy.

[0064] As described above, the present invention provides a method for precisely evaluating a patient's prognosis without invasive procedures by extracting quantified image information from CT images of a patient with the kidney disease and applying it to an artificial intelligence-based analysis model.

[0065] As used herein, the term “IgA nephropathy” refers to chronic glomerulonephritis caused by the abnormal deposition of immunoglobulin A (IgA) in the glomeruli. IgA nephropathy is known as the most common primary glomerulonephritis worldwide and causes hematuria, proteinuria, or renal decline. IgA nephropathy causes a progressive decline in renal function over a long period, leading to serious clinical problems that can progress to end-stage renal failure. More specifically, immunoglobulin A immune complexes are predominantly deposited in the glomerular mesangial region, causing mesangial cell proliferation and matrix increase, which may be accompanied by endothelial cell proliferation, segmental glomerular sclerosis, tubular atrophy, and interstitial fibrosis.

[0066]

[0067] According to a specific embodiment of the present invention, the aforementioned prognostic prediction value represents the percentage value of the likelihood that a patient with kidney disease will progress to end-stage renal disease (ESRD) within 5 years.

[0068] As used herein, the term “end-stage renal failure” refers to a condition in which the kidneys have suffered irreversible damage to the extent that they are unable to perform normal physiological functions, such as the excretion of waste products from the body, regulation of fluid and electrolyte balance, and hormone secretion. In the end-stage renal failure, the estimated glomerular filtration rate (eGFR) is significantly reduced, generally decreasing to less than 15 mL / min / 1.73 m², and patients require renal replacement therapy, such as dialysis or kidney transplantation, to sustain life due to the accumulation of toxic metabolites and fluid and electrolyte imbalances. End-stage renal failure is considered the final stage of progression for various chronic kidney diseases, including IgA nephropathy, diabetic nephropathy, lupus nephritis, and membranous nephropathy, and is used as a key evaluation indicator for predicting the patient’s long-term prognosis.

[0069] The present invention quantitatively calculates the probability that a patient will reach end-stage renal failure within 5 years by applying an image-based MEST score extracted from CT images of a patient with IgA nephropathy to a machine learning model. Accordingly, the present invention enables the repetitive and non-invasive evaluation of a patient's long-term prognosis without the need for an invasive kidney biopsy, which is the conventional method for calculating MEST scores.

[0070]

[0071] According to another aspect of the present invention, the present invention provides a prognosis prediction device for a patient with kidney disease comprising the following configuration:

[0072] (a) An extraction unit that extracts digitized image information from a segmented CT image of a patient with kidney disease;

[0073] (b) A feature selection unit that performs a feature selection process for the above image information;

[0074] (c) A score calculation unit that inputs the selected features into a machine learning model to calculate a MEST score;

[0075] (d) A prognosis prediction unit that applies the MEST score calculated by the machine learning above to a standard prognosis prediction tool.

[0076] The meanings of the MEST score, standard prognosis prediction tool, machine learning, feature selection process, kidney disease, end-stage renal failure, imaging information, etc. used in this invention are as described above, and are omitted to avoid excessive duplication.

[0077]

[0078] According to another aspect of the present invention, the present invention provides a method for training an artificial intelligence model for predicting the prognosis of a patient with kidney disease, comprising the following configuration:

[0079] (a) A step of extracting digitized image information from a CT image in which the lesion is segmented in a patient with kidney disease;

[0080] (b) a step of performing a feature selection process on the above image information;

[0081] (c) A step of training a machine learning model that predicts MEST (Mesangial hypercellularity, Endocapillary hypercellularity, Segmental glomerulosclerosis, Tubular atrophy / Interstitial fibrosis) scores using the above-mentioned selected features as input values.

[0082] The meanings of the MEST score, standard prognosis prediction tool, machine learning, feature selection process, kidney disease, end-stage renal failure, imaging information, etc. used in this invention are as described above, and are omitted to avoid excessive duplication.

[0083]

[0084] The features and advantages of the present invention are summarized as follows:

[0085] (a) The present invention provides a method for non-invasively calculating a MEST score that is directly related to the prognosis of a patient by extracting image features that have a close correlation with the prognosis of the patient from CT images of a patient with IgA nephropathy and training a machine learning model with these features.

[0086] (b) The present invention provides a non-invasive prognostic prediction technology that can replace or complement existing invasive and difficult-to-repeat kidney biopsies by rapidly predicting MEST scores and prognoses with high reliability based on CT images of patients with IgA nephropathy.

[0087] (c) By applying the image-based MEST score to a standard prognosis prediction tool, the present invention can quantitatively calculate the probability that a patient will progress to end-stage renal failure within a certain period, thereby providing important evidence for establishing personalized treatment strategies and long-term management plans for the patient.

[0088]

[0089] Figure 1 is a diagram illustrating patient data collected from two institutions and the processing thereof to develop an image-based prognosis prediction model for patients with IgA nephropathy. Figure 1a is a schematic diagram illustrating the process in which 261 patient data out of a total of 326 patient data collected from Institution 1 are classified into a training group and the remaining 65 patient data are classified into an internal validation group. Figure 1b is a schematic diagram illustrating the process in which a total of 41 patient data collected from Institution 2 are classified into an external validation group.

[0090] Figure 2 is a schematic diagram illustrating the process of segmenting a renal lesion region from a CT image using nnU-Net (neural network-based U-Net), extracting quantitative image features from the segmented renal cortex region using PyRadiomics (version 3.1.0) software, selecting features that have a significant correlation with prognosis prediction, and inputting them into a machine learning classifier.

[0091] FIG. 3 is a schematic diagram illustrating the operation process of a series of prognosis prediction devices provided by the present invention. It shows the process of extracting quantitative image features from segmented kidney regions and training them on a machine learning classifier to calculate a MEST score that has a significant correlation with the prognosis of a patient with kidney disease.

[0092] Figure 4 shows the performance of a machine learning model for calculating the mesangium proliferation (M) item among the aforementioned MEST scores. Figure 4a shows the Area Under the Receiver Operating Characteristic Curve (AUROC) of the model, Figure 4b shows the confusion matrix of the model, and Figure 4c shows the result of a feature importance analysis indicating the relative contribution of image features to the prediction of the mesangium proliferation item.

[0093] Figure 5 shows the performance of a machine learning model for calculating the endothelial cell proliferation (E) item among the aforementioned MEST scores. Figure 5a shows the Area Under the Receiver Operating Characteristic Curve (AUROC) of the model, Figure 5b shows the confusion matrix of the model, and Figure 5c shows the result of a feature importance analysis indicating the relative contribution of image features to the prediction of the endothelial cell proliferation item.

[0094] Figure 6 shows the performance of a machine learning model for calculating the segmental glomerular sclerosis (E) item among the aforementioned MEST scores. Figure 6a shows the Area Under the Receiver Operating Characteristic Curve (AUROC) of the model, Figure 6b shows the confusion matrix of the model, and Figure 6c shows the result of a feature importance analysis indicating the relative contribution of image features to the prediction of the segmental glomerular sclerosis item.

[0095] Figure 7 shows the performance of a machine learning model constructed to distinguish between T1 and T0·T2 for the tubular atrophy / interstitial fibrosis (T) item among the aforementioned MEST scores. Figure 7a shows the Area Under the Receiver Operating Characteristic Curve (AUROC) of the model, Figure 7b shows the confusion matrix of the model, and Figure 7c shows the result of a feature importance analysis indicating the relative contribution of image features to the prediction of the tubular atrophy / interstitial fibrosis item.

[0096] Figure 8 shows the performance of a machine learning model constructed to distinguish between T2 and T0·T1 for the tubular atrophy / interstitial fibrosis (T) item among the aforementioned MEST scores. Figure 8a shows the Area Under the Receiver Operating Characteristic Curve (AUROC) of the model, Figure 8b shows the confusion matrix of the model, and Figure 8c shows the result of a feature importance analysis indicating the relative contribution of image features to the prediction of the tubular atrophy / interstitial fibrosis item.

[0097] Figure 9 is a figure showing the results of regression analysis between Radiomics_MEST and Pathology_MEST for an internal validation group (n=65) and an external validation group (n=41) to verify the correlation between the Radiomics_MEST score calculated based on images and the Pathology_MEST score calculated by actual pathological readings.

[0098]

[0099] The present invention will be described in more detail below through examples. These examples are intended solely to explain the invention more specifically, and it will be obvious to those skilled in the art that the scope of the invention is not limited by these examples according to the gist of the invention.

[0100]

[0101] Examples

[0102] Patient Recruitment and Dataset Configuration

[0103] According to a specific embodiment of the present invention, patient data was collected from two institutions to develop an image-based prognosis prediction model for patients with IgA nephropathy.

[0104] (1) Data from 326 patients collected at Institution 1

[0105] A total of 383 patients were selected for whom IgA nephropathy was confirmed by biopsy between March 2010 and February 2023 and who underwent renal CT within one month of diagnosis. Of these, 57 patients meeting the following criteria were excluded from the analysis: cases where Oxford classification (MEST score) was not recorded (47 patients), cases where two biopsies were performed (1 patient), cases where pre-diagnosis CT images were not available (5 patients), cases where inappropriate imaging techniques were used (2 patients), and cases of kidney transplant patients (2 patients).

[0106] Therefore, a total of 326 patients were enrolled and randomly classified into a training cohort (n=261) and an internal validation cohort (n=65) (Fig. 1a).

[0107] (2) Data from 41 patients collected at Institution 2

[0108] Forty-nine patients diagnosed with IgA nephropathy by biopsy from March 2010 to February 2023 were selected by applying the same inclusion and exclusion criteria. Of these, a total of 41 patients were finally enrolled after excluding one patient who underwent two biopsies and seven patients for whom pre-diagnosis CT images were not available; these patients were used as the external validation cohort (Fig. 1b).

[0109] Accordingly, in this embodiment, the total number of patients used for the final analysis was 367, consisting of a training group and an internal validation group from Institution 1, and an external validation group from Institution 2. Through this multi-institution cohort-based dataset, image feature extraction, machine learning-based training, and independent external validation were performed.

[0110]

[0111] CT image preprocessing, feature extraction, and machine learning model construction

[0112] According to a specific embodiment of the present invention, a series of processes for calculating an image-based MEST score using CT images of a patient with IgA nephropathy is as follows.

[0113] First, all patients were analyzed using pre-contrast 3D CT images acquired at the time of diagnosis. Automatic segmentation was performed using nnU-Net (segmentation model) to separate the renal cortical region from the CT images, and the right and left kidneys were segmented independently.

[0114] Subsequently, PyRadiomics (version 3.1.0) software was used to extract quantitative radiomics features from the segmented renal cortex regions. The extracted features included various image-based indicators, including shape features, intensity features, and texture features.

[0115] To select key features that have a significant correlation with prognosis prediction from the extracted large set of features, the methods of Correlation (Corr), Relief, LASSO (Least Absolute Shrinkage and Selection Operator), PCA (Principal Component Analysis), and FFS (Forward Feature Selection) were applied.

[0116] The selected features were input into various machine learning classifiers to train the models. Specifically, a total of nine types of classifiers were applied, including DT (Decision Tree), RF (Random Forest), KNN (K-Nearest Neighbors), SGD (Stochastic Gradient Descent), SVM (Support Vector Machine), LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), AdaBoost, and GNB (Gaussian Naive Bayes). More than 63 models were constructed by varying the combination of each classifier and feature selection method, and hyperparameters were tuned using GridSearchCV for model optimization (Fig. 2).

[0117] The model was trained and validated by performing 5-fold cross-validation on a training cohort, and a classifier-feature selection method combination exhibiting optimal performance was derived. The model constructed in this invention is designed to independently predict the MEST scores of patients with IgA nephropathy, namely the M item (0 or 1), E item (0 or 1), S item (0 or 1), and T item (0, 1, or 2) (Fig. 3).

[0118]

[0119] Performance evaluation of machine learning models

[0120] According to a specific embodiment of the present invention, performance evaluations of various machine learning models constructed in Example 2 were performed on a training group, an internal validation group, and an external validation group. The performance of the prognosis prediction model according to the present invention was evaluated by calculating the Area Under the Receiver Operating Characteristic Curve (AUROC), TPR (sensitivity), TNR (specificity), PPV (precision), NPV (negative predictive value), FPR (false positive rate), FNR (false negative rate), FDR (false detection rate), ACC (accuracy), and confusion matrix of each model.

[0121] AUROC (Area Under the Receiver Operating Characteristic Curve) is an indicator representing the overall classification performance of the model; TPR (True Positive Rate) is the proportion of actual positives correctly predicted as positive by the model; TNR (True Negative Rate) is the proportion of actual negatives correctly predicted as negative; PPV (Positive Predictive Value) is the proportion of actual positives among samples predicted as positive; and NPV (Negative Predictive Value) is the proportion of actual negatives among samples predicted as negative. FPR (False Positive Rate) is the proportion of actual negatives incorrectly predicted as positive; FNR (False Negative Rate) is the proportion of actual positives incorrectly predicted as negative; FDR (False Discovery Rate) is the proportion of actual negatives among those predicted as positive; and ACC (Accuracy) is the proportion of correct predictions among the total samples.

[0122] In addition, the confusion matrix classifies prediction results into four categories by comparing them with actual values. TN (True Negative) refers to cases where a true negative is correctly classified as negative, and FP (False Positive) refers to cases where a true negative is incorrectly predicted as positive. FN (False Negative) refers to cases where a true positive is incorrectly classified as negative by the model, and TP (True Positive) refers to cases where a true positive is correctly predicted as positive.

[0123]

[0124] Prediction of the M item of the MEST score

[0125] According to a specific embodiment of the present invention, combinations of various feature selection techniques and classifiers were evaluated to predict whether mesangium proliferated (M0 and M1). As a result, the combination of Forward Feature Selection (FFS) and the AdaBoost classifier exhibited the best performance and was selected as the optimal model. In the training group, the model recorded an AUROC of 0.81, and in the internal validation group and external validation group, AUROCs of 0.68 and 0.76 were recorded, respectively, maintaining stable performance across various datasets (Fig. 4a).

[0126] The confusion matrix of the external validation group consisted of 26 TN cases, 6 FP cases, 4 FN cases, and 5 TP cases, and the accuracy calculated from this was 0.76 (Fig. 4b). More specifically, an analysis of the external validation group showed a sensitivity of 0.56, a specificity of 0.81, a precision of 0.45, and a negative predictive value of 0.87, with a false positive rate of 0.19, a false negative rate of 0.44, and a false detection rate of 0.55. In particular, through high specificity (0.81) and negative predictive value (0.87), it was experimentally confirmed that the model can reliably classify patients without mesangium proliferation.

[0127] Analysis of the internal validation group revealed that the model's accuracy was calculated as 0.70; while sensitivity was low at 0.38, specificity was 0.78 and negative predictive value was 0.84. In other words, although there was a tendency to miss some positive cases, the discriminative power for negative patients was consistently maintained, with a false positive rate of 0.23, a false negative rate of 0.62, and a false detection rate of 0.71. This suggests that the model's detection performance was somewhat limited in the internal validation group, but it still demonstrated clinically significant performance in terms of the stable classification of negative patients.

[0128] For the training group, the model showed balanced performance with an accuracy of 0.76, sensitivity of 0.72, specificity of 0.80, precision of 0.78, and negative predictive value of 0.71. The false positive rate was calculated as 0.20, the false negative rate as 0.28, and the false detection rate as 0.22, which means that the model demonstrated stable classification ability without overfitting on the training data. Therefore, the model of the present invention demonstrated a consistent level of performance not only in the training group but also in internal and external validation groups, and experimentally proved that it secured generalizability to various patient groups.

[0129] In addition, SHAP-based feature importance analysis revealed that radiomics features such as wavelet-HHL_gldm_DependenceEntropy, wavelet-LHL_glszm_SmallAreaLowGrayLevelEmphasis, and diagnostics_Mask-original_VoxelNum contributed significantly to the prediction of mesangium proliferation. This suggests that high-dimensional imaging features reflecting not only simple image intensity but also tissue structural and morphological information play a key role in prognosis prediction (Fig. 4c).

[0130] Therefore, the model of the present invention can determine the presence of mesangium proliferation with high specificity and reproducibility based on image features extracted from CT images, thereby demonstrating its potential as a non-invasive alternative technology capable of predicting a patient's pathological indicators without the need for an invasive kidney biopsy.

[0131]

[0132] Prediction of Item E of the MEST score

[0133] According to a specific embodiment of the present invention, various combinations were evaluated to predict whether endothelial cell proliferation occurred (E0 and E1), and as a result, a combination of Forward Feature Selection (FFS) and Decision Tree (DT) classifiers was derived as the optimal model. In the training group, the model showed very high performance with an AUROC of 0.99, and in the internal validation group and external validation group, AUROCs of 0.60 and 0.64 were calculated, respectively, maintaining a certain level of discriminative ability even in actual clinical application situations (Fig. 5a).

[0134] The confusion matrix generated from the external validation group consisted of 16 TN cases, 12 FP cases, 4 FN cases, and 9 TP cases, with an accuracy of 0.61 (Fig. 5b). More specifically, an analysis of the external validation group showed a sensitivity of 0.69, a specificity of 0.57, a precision of 0.43, and a negative predictive value of 0.80. The false positive rate was calculated to be 0.43, the false negative rate 0.31, and the false detection rate 0.57. This indicates that the model maintains a certain level of reliability in classifying negative patients, while the detection rate of positive patients is not excessively low. In particular, the negative predictive value (0.80) suggests that the model can be usefully utilized to exclude patients without endothelial cell proliferation.

[0135] In the internal validation group, the accuracy was 0.63, sensitivity 0.49, and specificity 0.69, while the precision was 0.36 and the negative predictive value was 0.79. The false positive rate was 0.31, the false negative rate was 0.51, and the false detection rate was 0.64. This indicates that the model's detection performance was somewhat limited in the internal validation group, but it still maintained a certain level of classification performance.

[0136] On the other hand, the training group recorded near-perfect values, including an AUROC of 0.99, accuracy of 0.99, sensitivity of 0.99, and specificity of 0.99. While this suggests the possibility that the FFS-DT model may overfit on the training data, the fact that it maintained an AUROC of 0.60-0.64 and accuracy of 0.61-0.63 in the inside and outside validation groups experimentally confirmed that its clinical applicability was secured.

[0137] As a result of feature importance analysis, it was confirmed that radiomic features such as wavelet-HLH_firstorder_Uniformity, wavelet-LHL_glszm_GrayLevelNonUniformity, and wavelet-LHH_glszm_ZoneVariance contribute significantly to distinguishing the presence of endothelial cell proliferation based on weight calculation using a Decision Tree classifier (DT) (Fig. 5c). This implies that imaging attributes reflecting cell density, uniformity, and regional dispersion within the tissue have a significant correlation with the presence of pathological endothelial cell proliferation. Therefore, the model of the present invention has been experimentally proven to serve as an image-based discrimination tool capable of non-invasively predicting endothelial cell proliferation, thereby complementing pathological diagnosis and making a significant contribution to the evaluation of patient prognosis.

[0138]

[0139] Prediction of the S item of the MEST score

[0140] According to a specific embodiment of the present invention, combinations of various feature selection techniques and classifiers were evaluated to predict the presence of segmental glomerular sclerosis (S0 and S1). As a result, the combination of Forward Feature Selection (FFS) and the AdaBoost classifier showed optimal performance. In the training group, the model recorded an AUROC of 0.90, while in the internal validation group and the external validation group, the AUROCs were 0.52 and 0.60, respectively, confirming that there is a difference in discriminative power depending on the dataset (Fig. 6a).

[0141] The confusion matrix generated from the external validation group consisted of 5 TN cases, 5 FP cases, 8 FN cases, and 23 TP cases, with an accuracy of 0.68 (Fig. 6b). More specifically, an analysis of the external validation group yielded a sensitivity of 0.74, a specificity of 0.50, a precision of 0.82, and a negative predictive value of 0.38. Additionally, the false positive rate was calculated as 0.50, the false negative rate as 0.26, and the false detection rate as 0.18. In particular, the high precision (0.82) indicates that there is a high probability that sclerosis actually exists among the cases identified as positive by the model, which implies that clinically reliable results are derived.

[0142] In the internal validation group, the accuracy was calculated as 0.59, sensitivity as 0.64, specificity as 0.41, precision as 0.79, and negative predictive value as 0.25, while the false positive rate was confirmed to be 0.59, the false negative rate as 0.36, and the false detection rate as 0.21. In other words, although there was a tendency for false positives to occur in some negative patients, the detection power (sensitivity 0.64) and precision (0.79) of positive patients were maintained at a certain level.

[0143] In the training group, balanced performance was demonstrated with accuracy 0.80, sensitivity 0.82, specificity 0.79, precision 0.79, and negative predictive value 0.81, while relatively stable results were shown with a false positive rate of 0.21, a false negative rate of 0.18, and a false detection rate of 0.21. This demonstrates that the model of the present invention maintained high detection performance for positive patients while minimizing overfitting.

[0144] As a result of the SHAP-based feature importance analysis, wavelet-HHL_glszm_LowGrayLevelZoneEmphasis showed the greatest contribution, and in addition, radiomics features such as exponential_firstorder_Variance, wavelet-HHH_glszm_ZoneVariance, and wavelet-HHL_firstorder_Skewness contributed significantly to the prediction of S item (Fig. 6c). These features are high-dimensional information reflecting the gray level distribution and variability within the image, and are interpreted as having a significant association with the actual pathological presence of segmental glomerular sclerosis. Therefore, the model of the present invention has experimentally proven that it can predict the presence of segmental glomerular sclerosis to a certain degree based on features extracted from CT images. In particular, since it has high precision and strengths in identifying benign patients, it can be utilized as an auxiliary decision-making tool in clinical practice.

[0145]

[0146] Prediction of the T-item of the MEST score

[0147] (1) Feature selection method and classifier to distinguish T0, T2 and T1

[0148] According to a specific embodiment of the present invention, various feature selection methods and classifiers were evaluated to distinguish between T0, T2, and T1 among tubular atrophy and interstitial fibrosis (T) items. As a result, the combination of the Least Absolute Shrinkage and Selection Operator (LASSO) and the Random Forest (RF) classifier showed optimal performance. In the training group, the model recorded an AUROC of 0.94, and in the internal validation group and external validation group, AUROCs of 0.70 and 0.76 were recorded, respectively, maintaining relatively stable discriminative power across various datasets (Fig. 7a).

[0149] The confusion matrix of the external validation group consisted of 35 TN cases, 3 FP cases, 1 FN case, and 2 TP cases, and the accuracy calculated from this was 0.90 (Fig. 7b). More specifically, an analysis of the external validation group showed that the sensitivity was 0.67, the specificity was 0.92, the precision was 0.40, and the negative predictive value was 0.97, while the false positive rate was calculated as 0.08, the false negative rate as 0.33, and the false detection rate as 0.60. In other words, the model has strengths in classifying negative patients, and in particular, the high specificity (0.92) and negative predictive value (0.97) mean that it can reliably exclude patients who are not clinically T1.

[0150] In the internal validation group, the accuracy was 0.75, sensitivity 0.50, and specificity 0.79, while the precision was 0.28 and the negative predictive value was 0.91. The false positive rate was calculated as 0.21, the false negative rate as 0.50, and the false detection rate as 0.72. This indicates that there were limitations in detecting some positive patients, but the ability to distinguish negative patients was maintained at a certain level.

[0151] In the training group, high performance was demonstrated with accuracy 0.87, sensitivity 0.93, specificity 0.82, precision 0.84, and negative predictive value 0.92, while the false positive rate was calculated as 0.18, the false negative rate as 0.07, and the false detection rate as 0.16. This suggests that the model demonstrated balanced performance on the training data.

[0152] As a result of RF-based Feature Importance analysis, features such as original_shape_Maximum2DDiameterColumn, exponential_gldm_DependenceNonUniformity, square_root_gldm_DependenceNonUniformity, and wavelet-LLL_glszm_DependenceNonUniformity showed high importance (Fig. 7c). These features serve as indicators reflecting structural non-uniformity and morphological characteristics within the image, and were found to be closely associated with tubular atrophy and interstitial fibrosis.

[0153] Therefore, the model of the present invention has demonstrated that it can differentiate T0, T2, and T1 with relatively high specificity and accuracy by utilizing CT image-based radiomics features, showing that it can function as a non-invasive auxiliary tool for predicting patient prognosis without the need for pathological examinations.

[0154]

[0155] (2) Feature selection method and classifier to distinguish T0, T1, and T2

[0156] According to a specific embodiment of the present invention, various combinations were evaluated to distinguish T0, T1, and T2 in the Tubular atrophy / Interstitial fibrosis (T) category. As a result, a combination of Forward Feature Selection (FFS) and Random Forest (RF) classifiers was selected as the optimal model. In the training group, the model recorded an AUROC of 0.98, and in the internal validation group and external validation group, the AUROCs were 0.73 and 0.93, respectively, confirming excellent discriminative power, particularly in the external validation group (Fig. 8a).

[0157] The confusion matrix of the external validation group consisted of 35 TNs, 5 FPs, 0 FNs, and 1 TP, with an accuracy of 0.88 (Fig. 8b). The sensitivity was 1.00, indicating that all actual T2 patients were detected; the specificity was 0.88, the precision was 0.17, and the negative predictive value was 1.00. The false positive rate was calculated as 0.12, the false negative rate as 0.00, and the false detection rate as 0.83. In other words, the model demonstrated the strength of not missing actual positive patients.

[0158] In the internal validation group, the accuracy was 0.88, sensitivity 0.25, specificity 0.91, precision 0.10, and negative predictive value 0.97. The false positive rate was 0.09, the false negative rate was 0.75, and the false detection rate was 0.90; while some positive detection performance was limited, the ability to distinguish negative patients was maintained.

[0159] In the training group, the model demonstrated very high performance with an accuracy of 0.95, sensitivity of 1.00, specificity of 0.91, precision of 0.92, and negative predictive value of 1.00, while the false positive rate was calculated as 0.09, the false negative rate as 0.00, and the false detection rate as 0.08. This indicates that the model achieved stable and balanced classification capabilities on the training data.

[0160] RF-based Feature Importance analysis results showed that image features such as wavelet-HLL_glszm_LowGrayLevelEmphasis, wavelet-HHL_firstorder_Skewness, and wavelet-HLL_glszm_LargeAreaHighGrayLevelEmphasis contributed significantly (Fig. 8c). These are high-dimensional indicators related to grayscale uniformity, skewness, and the presence of large high-gray-level areas within the image, suggesting that they are closely associated with pathological findings of tubular atrophy and interstitial fibrosis.

[0161] Therefore, the model of the present invention utilizes CT image-based radiomics features to distinguish between T0, T1, and T2 with high sensitivity, and possesses the potential to detect T2 patients without missing any. This has been experimentally proven to be a non-invasive prognostic prediction technology capable of predicting the progression stage of kidney disease and establishing personalized treatment strategies for patients without the need for pathological biopsy.

[0162]

[0163] Assessment of the correlation between radiomics-based MEST scores and pathological MEST scores

[0164] To verify the correlation between the Radiomics_MEST score calculated based on images and the Pathology_MEST score calculated from actual pathological readings, an analysis was performed on an internal validation group (n=65) and an external validation group (n=41).

[0165] The regression analysis results between Radiomics_MEST and Pathology_MEST in the internal validation group showed that the regression equation was y = 1.0536x, and the coefficient of determination (R²) was calculated to be 0.7877. This means that the Radiomics-based prediction score has a high correlation with actual pathological indicators, demonstrating that the image-based model of the present invention can stably reflect pathological findings (Fig. 9a).

[0166] In addition, regression analysis was performed between Radiomics_MEST(binary) and Pathology_MEST for the external validation group. As a result, the regression equation was derived as y = 1.0931x, and the R² was calculated as 0.8649. This means that the image-based MEST score shows a high degree of agreement with the pathology MEST score even in the external dataset, proving that the model of the present invention has secured generalizability between patient groups (Fig. 9b).

[0167] In addition, when the Radiomics_MEST score was applied to the International IgA Nephropathy Prognosis Prediction Tool (IIgAN-PT), it was experimentally confirmed that it can predict the risk of a patient's estimated glomerular filtration rate (eGFR) decreasing by more than 50% or progressing to end-stage renal disease (ESRD). Therefore, the image-based model of the present invention can be utilized as a useful clinical tool to precisely evaluate the long-term prognosis of patients by replacing or supplementing invasive kidney biopsy.

[0168]

[0169] Foregoing, specific parts of the present invention have been described in detail. It is evident to those skilled in the art that such specific descriptions are merely preferred embodiments and do not limit the scope of the present invention. Accordingly, the actual scope of the present invention is defined by the appended claims and their equivalents. The present invention will be described in more detail below through examples. These examples are intended solely to explain the present invention more specifically, and it will be obvious to those skilled in the art that the scope of the present invention is not limited by these examples according to the gist of the invention.

[0170]

[0171] Examples

[0172] Patient Recruitment and Dataset Configuration

[0173] According to a specific embodiment of the present invention, patient data was collected from two institutions to develop an image-based prognosis prediction model for patients with IgA nephropathy.

[0174] (1) Data from 326 patients collected at Institution 1

[0175] A total of 383 patients were selected for whom IgA nephropathy was confirmed by biopsy between March 2010 and February 2023 and who underwent renal CT within one month of diagnosis. Of these, 57 patients meeting the following criteria were excluded from the analysis: cases where Oxford classification (MEST score) was not recorded (47 patients), cases where two biopsies were performed (1 patient), cases where pre-diagnosis CT images were not available (5 patients), cases where inappropriate imaging techniques were used (2 patients), and cases of kidney transplant patients (2 patients).

[0176] Therefore, a total of 326 patients were enrolled and randomly classified into a training cohort (n=261) and an internal validation cohort (n=65) (Fig. 1a).

[0177] (2) Data from 41 patients collected at Institution 2

[0178] Forty-nine patients diagnosed with IgA nephropathy by biopsy from March 2010 to February 2023 were selected by applying the same inclusion and exclusion criteria. Of these, a total of 41 patients were finally enrolled after excluding one patient who underwent two biopsies and seven patients for whom pre-diagnosis CT images were not available; these patients were used as the external validation cohort (Fig. 1b).

[0179] Accordingly, in this embodiment, the total number of patients used for the final analysis was 367, consisting of a training group and an internal validation group from Institution 1, and an external validation group from Institution 2. Through this multi-institution cohort-based dataset, image feature extraction, machine learning-based training, and independent external validation were performed.

[0180]

[0181] CT image preprocessing, feature extraction, and machine learning model construction

[0182] According to a specific embodiment of the present invention, a series of processes for calculating an image-based MEST score using CT images of a patient with IgA nephropathy is as follows.

[0183] First, all patients were analyzed using pre-contrast 3D CT images acquired at the time of diagnosis. Automatic segmentation was performed using nnU-Net (segmentation model) to separate the renal cortical region from the CT images, and the right and left kidneys were segmented independently.

[0184] Subsequently, PyRadiomics (version 3.1.0) software was used to extract quantitative radiomics features from the segmented renal cortex regions. The extracted features included various image-based indicators, including shape features, intensity features, and texture features.

[0185] To select key features that have a significant correlation with prognosis prediction from the extracted large set of features, the methods of Correlation (Corr), Relief, LASSO (Least Absolute Shrinkage and Selection Operator), PCA (Principal Component Analysis), and FFS (Forward Feature Selection) were applied.

[0186] The selected features were input into various machine learning classifiers to train the models. Specifically, a total of nine types of classifiers were applied, including DT (Decision Tree), RF (Random Forest), KNN (K-Nearest Neighbors), SGD (Stochastic Gradient Descent), SVM (Support Vector Machine), LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), AdaBoost, and GNB (Gaussian Naive Bayes). More than 63 models were constructed by varying the combination of each classifier and feature selection method, and hyperparameters were tuned using GridSearchCV for model optimization (Fig. 2).

[0187] The model was trained and validated by performing 5-fold cross-validation on a training cohort, and a classifier-feature selection method combination exhibiting optimal performance was derived. The model constructed in this invention is designed to independently predict the MEST scores of patients with IgA nephropathy, namely the M item (0 or 1), E item (0 or 1), S item (0 or 1), and T item (0, 1, or 2) (Fig. 3).

[0188]

[0189] Performance evaluation of machine learning models

[0190] According to a specific embodiment of the present invention, performance evaluations of various machine learning models constructed in Example 2 were performed on a training group, an internal validation group, and an external validation group. The performance of the prognosis prediction model according to the present invention was evaluated by calculating the Area Under the Receiver Operating Characteristic Curve (AUROC), TPR (sensitivity), TNR (specificity), PPV (precision), NPV (negative predictive value), FPR (false positive rate), FNR (false negative rate), FDR (false detection rate), ACC (accuracy), and confusion matrix of each model.

[0191] AUROC (Area Under the Receiver Operating Characteristic Curve) is an indicator representing the overall classification performance of the model; TPR (True Positive Rate) is the proportion of actual positives correctly predicted as positive by the model; TNR (True Negative Rate) is the proportion of actual negatives correctly predicted as negative; PPV (Positive Predictive Value) is the proportion of actual positives among samples predicted as positive; and NPV (Negative Predictive Value) is the proportion of actual negatives among samples predicted as negative. FPR (False Positive Rate) is the proportion of actual negatives incorrectly predicted as positive; FNR (False Negative Rate) is the proportion of actual positives incorrectly predicted as negative; FDR (False Discovery Rate) is the proportion of actual negatives among those predicted as positive; and ACC (Accuracy) is the proportion of correct predictions among the total samples.

[0192] In addition, the confusion matrix classifies prediction results into four categories by comparing them with actual values. TN (True Negative) refers to cases where a true negative is correctly classified as negative, and FP (False Positive) refers to cases where a true negative is incorrectly predicted as positive. FN (False Negative) refers to cases where a true positive is incorrectly classified as negative by the model, and TP (True Positive) refers to cases where a true positive is correctly predicted as positive.

[0193]

[0194] Prediction of the M item of the MEST score

[0195] According to a specific embodiment of the present invention, combinations of various feature selection techniques and classifiers were evaluated to predict whether mesangium proliferated (M0 and M1). As a result, the combination of Forward Feature Selection (FFS) and the AdaBoost classifier exhibited the best performance and was selected as the optimal model. In the training group, the model recorded an AUROC of 0.81, and in the internal validation group and external validation group, AUROCs of 0.68 and 0.76 were recorded, respectively, maintaining stable performance across various datasets (Fig. 4a).

[0196] The confusion matrix of the external validation group consisted of 26 TN cases, 6 FP cases, 4 FN cases, and 5 TP cases, and the accuracy calculated from this was 0.76 (Fig. 4b). More specifically, an analysis of the external validation group showed a sensitivity of 0.56, a specificity of 0.81, a precision of 0.45, and a negative predictive value of 0.87, with a false positive rate of 0.19, a false negative rate of 0.44, and a false detection rate of 0.55. In particular, through high specificity (0.81) and negative predictive value (0.87), it was experimentally confirmed that the model can reliably classify patients without mesangium proliferation.

[0197] Analysis of the internal validation group revealed that the model's accuracy was calculated as 0.70; while sensitivity was low at 0.38, specificity was 0.78 and negative predictive value was 0.84. In other words, although there was a tendency to miss some positive cases, the discriminative power for negative patients was consistently maintained, with a false positive rate of 0.23, a false negative rate of 0.62, and a false detection rate of 0.71. This suggests that the model's detection performance was somewhat limited in the internal validation group, but it still demonstrated clinically significant performance in terms of the stable classification of negative patients.

[0198] For the training group, the model showed balanced performance with an accuracy of 0.76, sensitivity of 0.72, specificity of 0.80, precision of 0.78, and negative predictive value of 0.71. The false positive rate was calculated as 0.20, the false negative rate as 0.28, and the false detection rate as 0.22, which means that the model demonstrated stable classification ability without overfitting on the training data. Therefore, the model of the present invention demonstrated a consistent level of performance not only in the training group but also in internal and external validation groups, and experimentally proved that it secured generalizability to various patient groups.

[0199] In addition, SHAP-based feature importance analysis revealed that radiomics features such as wavelet-HHL_gldm_DependenceEntropy, wavelet-LHL_glszm_SmallAreaLowGrayLevelEmphasis, and diagnostics_Mask-original_VoxelNum contributed significantly to the prediction of mesangium proliferation. This suggests that high-dimensional imaging features reflecting not only simple image intensity but also tissue structural and morphological information play a key role in prognosis prediction (Fig. 4c).

[0200] Therefore, the model of the present invention can determine the presence of mesangium proliferation with high specificity and reproducibility based on image features extracted from CT images, thereby demonstrating its potential as a non-invasive alternative technology capable of predicting a patient's pathological indicators without the need for an invasive kidney biopsy.

[0201]

[0202] Prediction of Item E of the MEST score

[0203] According to a specific embodiment of the present invention, various combinations were evaluated to predict whether endothelial cell proliferation occurred (E0 and E1), and as a result, a combination of Forward Feature Selection (FFS) and Decision Tree (DT) classifiers was derived as the optimal model. In the training group, the model showed very high performance with an AUROC of 0.99, and in the internal validation group and external validation group, AUROCs of 0.60 and 0.64 were calculated, respectively, maintaining a certain level of discriminative ability even in actual clinical application situations (Fig. 5a).

[0204] The confusion matrix generated from the external validation group consisted of 16 TN cases, 12 FP cases, 4 FN cases, and 9 TP cases, with an accuracy of 0.61 (Fig. 5b). More specifically, an analysis of the external validation group showed a sensitivity of 0.69, a specificity of 0.57, a precision of 0.43, and a negative predictive value of 0.80. The false positive rate was calculated to be 0.43, the false negative rate 0.31, and the false detection rate 0.57. This indicates that the model maintains a certain level of reliability in classifying negative patients, while the detection rate of positive patients is not excessively low. In particular, the negative predictive value (0.80) suggests that the model can be usefully utilized to exclude patients without endothelial cell proliferation.

[0205] In the internal validation group, the accuracy was 0.63, sensitivity 0.49, and specificity 0.69, while the precision was 0.36 and the negative predictive value was 0.79. The false positive rate was 0.31, the false negative rate was 0.51, and the false detection rate was 0.64. This indicates that the model's detection performance was somewhat limited in the internal validation group, but it still maintained a certain level of classification performance.

[0206] On the other hand, the training group recorded near-perfect values, including an AUROC of 0.99, accuracy of 0.99, sensitivity of 0.99, and specificity of 0.99. While this suggests the possibility that the FFS-DT model may overfit on the training data, the fact that it maintained an AUROC of 0.60-0.64 and accuracy of 0.61-0.63 in the inside and outside validation groups experimentally confirmed that its clinical applicability was secured.

[0207] As a result of feature importance analysis, it was confirmed that radiomic features such as wavelet-HLH_firstorder_Uniformity, wavelet-LHL_glszm_GrayLevelNonUniformity, and wavelet-LHH_glszm_ZoneVariance contribute significantly to distinguishing the presence of endothelial cell proliferation based on weight calculation using a Decision Tree classifier (DT) (Fig. 5c). This implies that imaging attributes reflecting cell density, uniformity, and regional dispersion within the tissue have a significant correlation with the presence of pathological endothelial cell proliferation. Therefore, the model of the present invention has been experimentally proven to serve as an image-based discrimination tool capable of non-invasively predicting endothelial cell proliferation, thereby complementing pathological diagnosis and making a significant contribution to the evaluation of patient prognosis.

[0208]

[0209] Prediction of the S item of the MEST score

[0210] According to a specific embodiment of the present invention, combinations of various feature selection techniques and classifiers were evaluated to predict the presence of segmental glomerular sclerosis (S0 and S1). As a result, the combination of Forward Feature Selection (FFS) and the AdaBoost classifier showed optimal performance. In the training group, the model recorded an AUROC of 0.90, while in the internal validation group and the external validation group, the AUROCs were 0.52 and 0.60, respectively, confirming that there is a difference in discriminative power depending on the dataset (Fig. 6a).

[0211] The confusion matrix generated from the external validation group consisted of 5 TN cases, 5 FP cases, 8 FN cases, and 23 TP cases, with an accuracy of 0.68 (Fig. 6b). More specifically, an analysis of the external validation group yielded a sensitivity of 0.74, a specificity of 0.50, a precision of 0.82, and a negative predictive value of 0.38. Additionally, the false positive rate was calculated as 0.50, the false negative rate as 0.26, and the false detection rate as 0.18. In particular, the high precision (0.82) indicates that there is a high probability that sclerosis actually exists among the cases identified as positive by the model, which implies that clinically reliable results are derived.

[0212] In the internal validation group, the accuracy was calculated as 0.59, sensitivity as 0.64, specificity as 0.41, precision as 0.79, and negative predictive value as 0.25, while the false positive rate was confirmed to be 0.59, the false negative rate as 0.36, and the false detection rate as 0.21. In other words, although there was a tendency for false positives to occur in some negative patients, the detection power (sensitivity 0.64) and precision (0.79) of positive patients were maintained at a certain level.

[0213] In the training group, balanced performance was demonstrated with accuracy 0.80, sensitivity 0.82, specificity 0.79, precision 0.79, and negative predictive value 0.81, while relatively stable results were shown with a false positive rate of 0.21, a false negative rate of 0.18, and a false detection rate of 0.21. This demonstrates that the model of the present invention maintained high detection performance for positive patients while minimizing overfitting.

[0214] As a result of the SHAP-based feature importance analysis, wavelet-HHL_glszm_LowGrayLevelZoneEmphasis showed the greatest contribution, and in addition, radiomics features such as exponential_firstorder_Variance, wavelet-HHH_glszm_ZoneVariance, and wavelet-HHL_firstorder_Skewness contributed significantly to the prediction of S item (Fig. 6c). These features are high-dimensional information reflecting the gray level distribution and variability within the image, and are interpreted as having a significant association with the actual pathological presence of segmental glomerular sclerosis. Therefore, the model of the present invention has experimentally proven that it can predict the presence of segmental glomerular sclerosis to a certain degree based on features extracted from CT images. In particular, since it has high precision and strengths in identifying benign patients, it can be utilized as an auxiliary decision-making tool in clinical practice.

[0215]

[0216] Prediction of the T-item of the MEST score

[0217] (1) Feature selection method and classifier to distinguish T0, T2 and T1

[0218] According to a specific embodiment of the present invention, various feature selection methods and classifiers were evaluated to distinguish between T0, T2, and T1 among tubular atrophy and interstitial fibrosis (T) items. As a result, the combination of the Least Absolute Shrinkage and Selection Operator (LASSO) and the Random Forest (RF) classifier showed optimal performance. In the training group, the model recorded an AUROC of 0.94, and in the internal validation group and external validation group, AUROCs of 0.70 and 0.76 were recorded, respectively, maintaining relatively stable discriminative power across various datasets (Fig. 7a).

[0219] The confusion matrix of the external validation group consisted of 35 TN cases, 3 FP cases, 1 FN case, and 2 TP cases, and the accuracy calculated from this was 0.90 (Fig. 7b). More specifically, an analysis of the external validation group showed that the sensitivity was 0.67, the specificity was 0.92, the precision was 0.40, and the negative predictive value was 0.97, while the false positive rate was calculated as 0.08, the false negative rate as 0.33, and the false detection rate as 0.60. In other words, the model has strengths in classifying negative patients, and in particular, the high specificity (0.92) and negative predictive value (0.97) mean that it can reliably exclude patients who are not clinically T1.

[0220] In the internal validation group, the accuracy was 0.75, sensitivity 0.50, and specificity 0.79, while the precision was 0.28 and the negative predictive value was 0.91. The false positive rate was calculated as 0.21, the false negative rate as 0.50, and the false detection rate as 0.72. This indicates that there were limitations in detecting some positive patients, but the ability to distinguish negative patients was maintained at a certain level.

[0221] In the training group, high performance was demonstrated with accuracy 0.87, sensitivity 0.93, specificity 0.82, precision 0.84, and negative predictive value 0.92, while the false positive rate was calculated as 0.18, the false negative rate as 0.07, and the false detection rate as 0.16. This suggests that the model demonstrated balanced performance on the training data.

[0222] As a result of RF-based Feature Importance analysis, features such as original_shape_Maximum2DDiameterColumn, exponential_gldm_DependenceNonUniformity, square_root_gldm_DependenceNonUniformity, and wavelet-LLL_glszm_DependenceNonUniformity showed high importance (Fig. 7c). These features serve as indicators reflecting structural non-uniformity and morphological characteristics within the image, and were found to be closely associated with tubular atrophy and interstitial fibrosis.

[0223] Therefore, the model of the present invention has demonstrated that it can differentiate T0, T2, and T1 with relatively high specificity and accuracy by utilizing CT image-based radiomics features, showing that it can function as a non-invasive auxiliary tool for predicting patient prognosis without the need for pathological examinations.

[0224]

[0225] (2) Feature selection method and classifier to distinguish T0, T1, and T2

[0226] According to a specific embodiment of the present invention, various combinations were evaluated to distinguish T0, T1, and T2 in the Tubular atrophy / Interstitial fibrosis (T) category. As a result, a combination of Forward Feature Selection (FFS) and Random Forest (RF) classifiers was selected as the optimal model. In the training group, the model recorded an AUROC of 0.98, and in the internal validation group and external validation group, the AUROCs were 0.73 and 0.93, respectively, confirming excellent discriminative power, particularly in the external validation group (Fig. 8a).

[0227] The confusion matrix of the external validation group consisted of 35 TNs, 5 FPs, 0 FNs, and 1 TP, with an accuracy of 0.88 (Fig. 8b). The sensitivity was 1.00, indicating that all actual T2 patients were detected; the specificity was 0.88, the precision was 0.17, and the negative predictive value was 1.00. The false positive rate was calculated as 0.12, the false negative rate as 0.00, and the false detection rate as 0.83. In other words, the model demonstrated the strength of not missing actual positive patients.

[0228] In the internal validation group, the accuracy was 0.88, sensitivity 0.25, specificity 0.91, precision 0.10, and negative predictive value 0.97. The false positive rate was 0.09, the false negative rate was 0.75, and the false detection rate was 0.90; while some positive detection performance was limited, the ability to distinguish negative patients was maintained.

[0229] In the training group, the model demonstrated very high performance with an accuracy of 0.95, sensitivity of 1.00, specificity of 0.91, precision of 0.92, and negative predictive value of 1.00, while the false positive rate was calculated as 0.09, the false negative rate as 0.00, and the false detection rate as 0.08. This indicates that the model achieved stable and balanced classification capabilities on the training data.

[0230] RF-based Feature Importance analysis results showed that image features such as wavelet-HLL_glszm_LowGrayLevelEmphasis, wavelet-HHL_firstorder_Skewness, and wavelet-HLL_glszm_LargeAreaHighGrayLevelEmphasis contributed significantly (Fig. 8c). These are high-dimensional indicators related to grayscale uniformity, skewness, and the presence of large high-gray-level areas within the image, suggesting that they are closely associated with pathological findings of tubular atrophy and interstitial fibrosis.

[0231] Therefore, the model of the present invention utilizes CT image-based radiomics features to distinguish between T0, T1, and T2 with high sensitivity, and possesses the potential to detect T2 patients without missing any. This has been experimentally proven to be a non-invasive prognostic prediction technology capable of predicting the progression stage of kidney disease and establishing personalized treatment strategies for patients without the need for pathological biopsy.

[0232]

[0233] Assessment of the correlation between radiomics-based MEST scores and pathological MEST scores

[0234] To verify the correlation between the Radiomics_MEST score calculated based on images and the Pathology_MEST score calculated from actual pathological readings, an analysis was performed on an internal validation group (n=65) and an external validation group (n=41).

[0235] The regression analysis results between Radiomics_MEST and Pathology_MEST in the internal validation group showed that the regression equation was y = 1.0536x, and the coefficient of determination (R²) was calculated to be 0.7877. This means that the Radiomics-based prediction score has a high correlation with actual pathological indicators, demonstrating that the image-based model of the present invention can stably reflect pathological findings (Fig. 9a).

[0236] In addition, regression analysis was performed between Radiomics_MEST(binary) and Pathology_MEST for the external validation group. As a result, the regression equation was derived as y = 1.0931x, and the R² was calculated as 0.8649. This means that the image-based MEST score shows a high degree of agreement with the pathology MEST score even in the external dataset, proving that the model of the present invention has secured generalizability between patient groups (Fig. 9b).

[0237] In addition, when the Radiomics_MEST score was applied to the International IgA Nephropathy Prognosis Prediction Tool (IIgAN-PT), it was experimentally confirmed that it can predict the risk of a patient's estimated glomerular filtration rate (eGFR) decreasing by more than 50% or progressing to end-stage renal disease (ESRD). Therefore, the image-based model of the present invention can be utilized as a useful clinical tool to precisely evaluate the long-term prognosis of patients by replacing or supplementing invasive kidney biopsy.

[0238]

[0239] Foregoing, specific parts of the present invention have been described in detail. It is evident to those skilled in the art that such specific descriptions are merely preferred embodiments and do not limit the scope of the invention. Accordingly, the actual scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for providing information necessary for predicting the prognosis of patients with kidney disease, comprising the following steps: (a) A step of extracting digitized image information from a CT image in which the lesion is segmented in a patient with kidney disease; (b) a step of performing a feature selection process on the above image information; (c) a step of inputting the selected features into a machine learning model to calculate a MEST (Mesangial hypercellularity, Endocapillary hypercellularity, Segmental glomerulosclerosis, Tubular atrophy / Interstitial fibrosis) score; and (d) A step of obtaining a prognostic prediction value by inputting the MEST score calculated by the machine learning above into a standard prognostic prediction tool.

2. A method according to claim 1, wherein the feature selection process of step (b) is performed by the No FS (NO Feature Selection), Corr (Correlation-based Selection), LASSO (Least Absolute Shrinkage and Selection Operator), or FFS (Forward Feature Selection) method.

3. A method according to claim 2, wherein the feature selection process of step (b) is performed by the LASSO (Least Absolute Shrinkage and Selection Operator) or FFS (Forward Feature Selection) method.

4. A method according to claim 1, wherein the step of inputting the selected features of step (c) into a machine learning model is characterized by being selected by a binary classifier selected from a group consisting of Random Forest, Adaptive Boosting, KNN (K-Nearest Neighbors), SGD (Stochastic Gradient Descent), SVM (Support Vector Machine), LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), GNB (Gaussian Naive Bayes), SHAP (SHapley Additive exPlanations), and Decision Tree.

5. A method according to claim 4, wherein the step of inputting the selected features of step (c) into a machine learning model is performed by a Random Forest, SHAP (SHapley Additive exPlanations), or a Decision Tree.

6. A method according to claim 1, wherein the kidney disease is selected from the group consisting of IgA nephropathy, membranous nephropathy, minimal change nephropathy, diabetic nephropathy, lupus nephritis, polycystic nephropathy, and hypertensive nephropathy.

7. A method according to claim 1, characterized in that the kidney disease is IgA nephropathy.

8. A method according to claim 1, characterized in that the predicted prognosis value is a percentage value of the likelihood that the patient with the kidney disease will progress to end-stage renal disease (ESRD) within 5 years.

9. A prognosis prediction device for patients with kidney disease comprising the following configuration. (a) An extraction unit that extracts digitized image information from a segmented CT image of a patient with kidney disease; (b) A feature selection unit that performs a feature selection process for the above image information; (c) A score calculation unit that inputs the above-mentioned selected features into a machine learning model to calculate a MEST (Mesangial hypercellularity; M, Endocapillary hypercellularity; E, Segmental glomerulosclerosis; S, and Tubular atrophy / Interstitial fibrosis; T) score; (d) a prognosis prediction unit that applies the MEST score calculated by the machine learning above to a standard prognosis prediction tool; 10. An apparatus according to claim 9, characterized in that the feature selection process is performed by LASSO (Least Absolute Shrinkage and Selection Operator) or FFS (Forward Feature Selection).

11. An apparatus according to claim 9, wherein the machine learning model is performed by a group consisting of Random Forest, Adaptive Boosting, KNN (K-Nearest Neighbors), SGD (Stochastic Gradient Descent), SVM (Support Vector Machine), LDA (Linear Discriminant Analysis), QDA (Quadratic Discriminant Analysis), GNB (Gaussian Naive Bayes), SHAP (SHapley Additive exPlanations), and Decision Tree.

12. An apparatus according to claim 11, characterized in that the machine learning model is a Random Forest, SHAP (SHapley Additive exPlanations), or a Decision Tree.

13. A device according to claim 9, characterized in that the kidney disease is IgA nephropathy.

14. An apparatus according to claim 9, wherein the kidney disease is selected from the group consisting of IgA nephropathy, membranous nephropathy, minimal change nephropathy, diabetic nephropathy, lupus nephritis, polycystic nephropathy, and hypertensive nephropathy.

15. A device according to claim 9, characterized in that the prognosis prediction is a percentage value of the probability that the patient with the kidney disease will progress to end-stage renal disease (ESRD) within 5 years.

16. A training method for an artificial intelligence model for predicting the prognosis of patients with kidney disease, comprising the following steps: (a) A step of extracting digitized image information from a CT image in which the lesion is segmented in a patient with kidney disease; (b) a step of performing a feature selection process on the above image information; (c) A step of training a machine learning model that predicts MEST (Mesangial hypercellularity, Endocapillary hypercellularity, Segmental glomerulosclerosis, Tubular atrophy / Interstitial fibrosis) scores using the above-mentioned selected features as input values.

17. A method according to claim 16, characterized in that the kidney disease is IgA nephropathy.

18. A method according to claim 16, wherein the prognosis prediction is a percentage value of the likelihood that the patient with the kidney disease will progress to end-stage renal disease (ESRD) within 5 years.

Images