Elderly chronic kidney disease progress prediction model construction method based on kidney phenotype map

Abstract

The invention discloses a kidney phenotype map-based elderly chronic kidney disease progress prediction model construction method, and belongs to the technical field of disease prediction. The method comprises the following steps: S1, determining a selection standard and an exclusion standard, and defining a research population; determining a research type and a blind method design; s2, four phenotype databases are constructed in a standardized mode through the whole process of data acquisition, processing and storage, and multi-dimensional data from macroscopic clinical information to microscopic molecular markers are covered; s3, making a follow-up plan and defining an end point; s4, determining the minimum sample size based on double logics of risk factor screening and the outcome occurrence rate; s5, constructing a model by adopting a classical machine learning algorithm, and evaluating the performance of the model through a multi-dimensional index; constructing a model of various different phenotype combinations; and S6, model comparison and screening: screening an optimal clinical application model from the constructed various models through specificity and sensitivity comparison. According to the invention, a multi-dimensional, high-precision and landing prediction model can be obtained.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of disease prediction, and particularly relates to a method for constructing an old chronic kidney disease progression prediction model based on a kidney phenotype atlas by using an artificial intelligence method. BACKGROUND

[0002] Old chronic kidney disease (CKD) is a core factor leading to old uremia and health risk multiplication due to high incidence and concealment of multiple organ function decline, multiple disease comorbidities, and reduced kidney function. Globally, there are more than 697 million CKD patients, and the related mortality rate has increased by 53% in the same period. Aging is the main reason for the increase in the prevalence of CKD and the progression of kidney function. With age, the average annual decline of GFR in adults is 0.75-1.0 mL / min / 1.73 m², and about 6000-6500 kidney units are lost per year after the age of 30. The decline of GFR in the elderly is divided into two categories: "healthy aging" and "disease acceleration", and the prognosis is very different and needs to be distinguished.

[0003] Current CKD management relies on risk prediction models, but existing models are mostly based on electronic medical record data (demographics, routine examinations), such as the multi-country cohort model published in JAMA in 2019. The model is based on eGFR<45 mL / min / 1.73 m 2 Patient calibration is poor; at the same time, there is a lack of research on kidney phenotype groups (molecules, functions, images, etc.), traditional biomarkers (creatinine / eGFR, urinary albumin) have low sensitivity and large fluctuations, and there is a lack of non-invasive detection technology for kidney fibrosis on images, which greatly limits the understanding of the progression of old CKD.

[0004] However, the existing technology has the following main deficiencies: 1. Single dimension of prediction model: only relying on electronic medical record data, without integrating multi-dimensional phenotypes such as molecules, images, etc., which cannot cover the complexity of old CKD, resulting in poor prediction accuracy for some patients.

[0005] 2. Incomplete kidney phenotype group: the phenotypes such as molecules (new biomarkers), functions (renal tubular function), and images (multi-modal MRI) have not been systematically constructed, and the correlation between phenotypes is not clear, which cannot fully analyze the kidney status.

[0006] 3. Defects of biomarkers and imaging technology: traditional markers are difficult to identify CKD at an early stage, and there is a lack of non-invasive kidney fibrosis evaluation methods.

[0007] Therefore, the existing CKD prediction technology cannot meet the high requirements of the current and future fields for CKD prediction, and how to improve the early accurate assessment and identification of CKD has become a key point. SUMMARY

[0008] The present application aims to at least partially solve one of the above-mentioned technical problems in the related art.

[0009] To this end, the present application aims to provide a kidney phenotype atlas-based elderly chronic kidney disease progression prediction model construction method, which can specifically address the core pain points of existing elderly chronic kidney disease (CKD) prediction technologies and obtain a multi-dimensional, high-precision, and landable prediction model.

[0010] To solve the above technical problems, the present application is implemented as follows: The present application embodiment provides a kidney phenotype atlas-based elderly chronic kidney disease progression prediction model construction method, which comprises: S1. Basic design: determine the inclusion criteria and exclusion criteria, define the study population; determine the research type and blind design; S2. Four-dimensional kidney phenotype construction: through data acquisition-processing-storage full-process standardization, construct four phenotype databases of clinical, molecular, functional, and imaging, covering multi-dimensional data from macro clinical information to micro molecular markers; S3. Formulate a follow-up plan and define endpoints; S4. Sample size calculation: based on the double logic of risk factor screening and outcome incidence, determine the minimum sample size; S5. AI prediction model construction and verification: use classic machine learning algorithms to construct the model, evaluate the model performance through multi-dimensional indicators, and ensure the prediction accuracy; construct models of various different phenotype combinations; S6. Model comparison and screening: compare the specificity and sensitivity to screen the optimal clinical application model from the constructed multiple models.

[0011] In addition, the kidney phenotype atlas-based elderly chronic kidney disease progression prediction model construction method according to the present application can also have the following additional technical features: In some embodiments thereof, the four-dimensional kidney phenotype of step S2 comprises: a kidney clinical phenotype, a kidney molecular phenotype, a kidney functional phenotype, and a kidney imaging phenotype.

[0012] In some embodiments thereof, the data content of the kidney clinical phenotype comprises: Basic clinical indicators: gender, age, height, weight, BMI, systolic blood pressure, diastolic blood pressure, fasting blood glucose, glycosylated hemoglobin, cholesterol, triglyceride, low-density lipoprotein, and high-density lipoprotein; Comorbidities: heart failure, coronary heart disease, arrhythmia, myocardial infarction, hypertension, hyperlipidemia, peripheral atherosclerosis, cerebral hemorrhage, cerebral infarction, chronic obstructive pulmonary disease, asthma, hyperuricemia, gout, metabolic acidosis, and hyperkalemia; Personal history: smoking history and drinking history; Concurrent medication: insulin, metformin, sulfonylureas, renin-angiotensin-aldosterone system inhibitors, mineralocorticoid receptor antagonists, ARB, CCB, diuretics, beta blockers, non-steroidal anti-inflammatory drugs and calcineurin inhibitors; The core biomarkers of the renal molecular phenotype include co-targeting 16 biomarkers related to CKD progression; specifically including: Primary biomarkers: tumor necrosis factor receptor 1, tumor necrosis factor receptor 2, kidney injury molecule 1 and neutrophil gelatinase-associated lipocalin; Secondary biomarkers: cystatin C, asymmetric dimethylarginine, liver-type fatty acid-binding protein, tissue inhibitor of metalloproteinase 2, insulin-like growth factor binding protein 7, alpha 1-microglobulin, urinary albumin, interleukin 18, monocyte chemotactic protein 1, chitinase 3-like protein 1, type III procollagen amino-terminal peptide and epidermal growth factor; The data of the renal function phenotype include: Glomerular function indicators: serum creatinine, serum Cystatin C, urinary total protein / creatinine, urinary albumin / creatinine, 24-hour urinary protein quantification; Tubular function indicators: urinary NAG / creatinine, urinary beta 2-microglobulin, urinary alpha 1-microglobulin, urinary transferrin, urinary free light chain kappa and lambda; The renal imaging phenotype: based on multi-modal MRI technology, non-invasive and quantitative renal imaging data is constructed, and the pathophysiological state including renal fibrosis and blood perfusion is evaluated.

[0013] In some embodiments, the content of the follow-up plan in step S3 includes: the follow-up period is 1 year, the number of effective serum creatinine measurements within 1 year is required to be ≥2 times, and the change trend of eGFR is ensured to be traceable; all detection items in the baseline period are repeated at the end of the follow-up period; for patients who reach the renal composite endpoint, repeat the renal multi-modal MRI examination to evaluate the changes in the imaging phenotype; The endpoint is the renal composite endpoint, and eGFR is calculated by CKD-EPI formula. If any of the following conditions is met, it is determined that the endpoint is reached: Acute renal function decline: eGFR slope decreases by ≥5 ml / min / 1.73m 2 / year; eGFR absolute value decreases by ≥40% of the baseline value; Entering end-stage renal disease: eGFR ≤15 ml / min / 1.73m 2 .

[0014] In some embodiments, the sample size calculation in step S4 includes: estimating the number of selected variables of the model according to the general principles of risk factor screening design, taking 10-20 times of the number of selected variables as the standard, dividing by the incidence of eGFR slope decline ≥5ml / min / 1.73m 2 2 in elderly CKD patients per year, and considering the missing information, calculating the value of the sample size.

[0015] In some embodiments, the step of constructing an AI prediction model in step S5 includes: S51. Establish a standardized database: standardize the clinical, molecular, functional, and imaging four-dimensional phenotype data, associate the unique identifier of the patient, and form a structured CKD database for subsequent modeling; S52. Data preprocessing: clean the data and perform data discretization; S53. Feature selection: using machine learning algorithms, taking four-dimensional phenotype variables as input and kidney composite endpoints as output, screening features significantly related to CKD progression, and excluding irrelevant variables; S54. Model training: divide the obtained feature data into training set, validation set and test set in the ratio of 7:1:2, train, parameter optimize and adjust the classic machine learning algorithm model, and evaluate the generalization ability.

[0016] In some embodiments, the classic machine learning algorithm includes multilayer perceptron, support vector machine, random forest and Adaboost algorithm, etc., and 1-2 methods are selected.

[0017] In some embodiments, the validation in step S5 is to evaluate the model performance by multi-dimensional indicators to ensure prediction accuracy; the multi-dimensional indicators specifically include: Accuracy: the number of correctly classified samples / total number of samples, reflecting the overall prediction ability; Sensitivity: true positive sample number / (true positive + false negative), reflecting the recognition ability of patients progressing to endpoints; Specificity: true negative sample number / (true negative + false positive), reflecting the recognition ability of patients not progressing; Precision: true positive sample number / (true positive + false positive), reflecting the reliability of positive prediction results; ROC curve: the curve is drawn with false positive rate as horizontal axis and true positive rate as vertical axis, and the larger the AUC value, the stronger the model discrimination ability.

[0018] In some embodiments, step S5 constructs models of four different phenotype combinations, including: Model 1: clinical phenotype + functional phenotype, basic combination, covering traditional clinical indicators; Model 2: functional phenotype + molecular phenotype, focus on function + molecule, strengthen micro-mechanism correlation; Model 3: clinical phenotype + molecular phenotype + imaging phenotype, exclude functional phenotype, verify multi-modal synergy; and, Model 4: clinical phenotype + functional phenotype + molecular phenotype + imaging phenotype, full-dimensional combination.

[0019] In some embodiments, the content of step S6 includes: comparing the verification indexes of the four models through the same queue data verification, and selecting the model with the optimal performance as the old CKD progression prediction model for clinical practice.

[0020] Compared with the prior art, the present application has at least the following beneficial effects: In the embodiment of the present application, the provided old chronic kidney disease progression prediction model based on kidney phenotype atlas constructs a "clinical phenotype + molecular phenotype + functional phenotype + imaging phenotype" four-dimensional database, covering full-scale information from micro-molecules to macro-organ, breaking through the traditional mode of "only relying on electronic medical record data"; The clinical phenotype adds subdivision variables such as combined diseases (ICD9 / 10 code), combined drugs (such as RAASi, MRA), etc., instead of only basic demographic data; The molecular phenotype targets the screening of 16 biomarkers (including 4 main markers such as TNFR1 and KIM-1), adopts "high-throughput mass spectrometry + electrochemiluminescence" double quantitative method, and replaces the traditional single creatinine / eGFR index; The functional phenotype supplements the evaluation of tubular function (urinary NAG / creatinine, urinary beta2MG, etc.), filling the gap of only focusing on glomerular function; The imaging phenotype adopts multi-modal MRI (BOLD, IVIM, DKI, PWI sequence), non-invasive assessment of renal hypoxia, microcirculation and fibrosis, replacing the current situation of non-invasive imaging technology; The four structures are different and complementary; The molecular phenotype captures early renal injury signals (such as TNFR1 reflecting inflammation), the imaging phenotype visualizes the changes of renal structure and function (such as DKI evaluating fibrosis), combined with clinical / functional data, which can present the kidney status of old CKD patients in a panoramic view, avoiding the problem of "missing early lesions only relying on clinical data"; The four-dimensional data are complementary - for example, the clinical phenotype reflects the risk of underlying diseases, the molecular phenotype suggests early damage, the imaging phenotype verifies pathological changes, and joint modeling can cover the "risk-damage-progression" whole chain, solving the problem of eGFR<45mL / min / 1.73m 2 patient calibration problems; In the embodiment of the present application, the kidney phenotype-based elderly chronic kidney disease progression prediction model construction method provided formulates precise admission and exclusion criteria (such as only including CKD 3-4 stage patients over 60 years old), excludes interference factors such as acute kidney injury and kidney transplantation, ensures that the research population is consistent with the real clinical high-risk group of elderly CKD, and standardizes follow-up and endpoint definition to solve the problems of incomplete follow-up data and ambiguous endpoints. In the embodiment of the present application, the kidney phenotype-based elderly chronic kidney disease progression prediction model construction method provided is based on the CKD diagnosis and treatment method innovation of artificial intelligence model explainability analysis: using the explainability deep learning frontier method, the correlation between each modality magnetic resonance functional imaging and kidney function indicators, clinical indicators and molecular markers, and the importance of CKD progression and prognosis prediction are quantified, and the elderly CKD diagnosis and treatment method is optimized.

[0021] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The prediction model construction flowchart disclosed for an embodiment of the present application; Figure 2 The technical roadmap disclosed for embodiment 1 of the present application; Figure 3 The ROC curve of the clinical + functional + biomarker + magnetic resonance data set in the random forest model disclosed for embodiment 1 of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] The embodiments of the present application will be described in detail below with reference to the drawings and specific embodiments and their application scenarios.

[0025] In some embodiments of the present application, a method for predicting the progression of chronic kidney disease in the elderly based on artificial intelligence is provided. In a large sample prospective cohort of elderly patients with CKD, electronic medical record clinical data variables are collected to form a kidney clinical phenotype database, a blood and urine specimen bank is established to screen for molecular biomarkers to construct a kidney molecular phenotype, functional magnetic resonance imaging technology is used to construct a kidney imaging phenotype, and special test items and methods are used to establish a kidney function phenotype. Based on the four kidney phenotypes, artificial intelligence methods are used to analyze the kidney phenotype group atlas, determine the correlation between the phenotypes, and create a CKD prediction model to predict the progression and prognosis of kidney function in elderly patients with CKD.

[0026] The model construction steps include: Step 1. Establish a large sample prospective cohort of elderly patients with chronic kidney disease and a blood and urine specimen bank (including baseline and follow-up), and screen for molecular biomarkers that can predict the progression of CKD; Step 2. Establish a clinical phenotype database for a large sample prospective cohort of elderly patients with chronic kidney disease (including baseline and follow-up); Step 3. Establish a magnetic resonance imaging phenotype database for a large sample prospective cohort of elderly patients with chronic kidney disease (including baseline and follow-up); Step 4. Establish a functional phenotype database for a large sample prospective cohort of elderly patients with chronic kidney disease (including baseline and follow-up); Step 5. Apply artificial intelligence methods to integrate kidney phenotype groups, construct a phenotype group atlas, determine the correlation between kidney phenotypes, and establish a CKD progression prediction model.

[0027] Drawing a kidney phenotype group atlas: Kidney disease is one of the major chronic diseases that affect population health. Due to the lack of understanding of its molecular mechanisms, its diagnosis is still at the level of clinical syndromes and pathological morphology. Conducting clinical research based on precise phenotypes and constructing a precise disease diagnosis and classification system is an important breakthrough in improving the level of kidney disease diagnosis and treatment and research capabilities. The present application applies phenotype group research methods to kidney disease research based on clinical needs. The molecular phenotype obtained from the research of new biomarkers of the kidney, the imaging phenotype of the multi-modal model of kidney function magnetic resonance imaging, and the common clinical influencing factors screening (clinical phenotype) and kidney function assessment results (functional phenotype) are combined to realize four-dimensional kidney phenotype group panoramic analysis and truly realize the drawing of a kidney phenotype group atlas.

[0028] Construction of kidney imaging phenotype group: functional magnetic resonance imaging can quantitatively analyze the renal microcirculation blood perfusion state and renal filtration function, and its diagnostic value in the evaluation of early renal function damage of CKD is increasingly prominent, and has become an imaging research hotspot for renal function evaluation. The inventors have clarified the relationship between the effective sequence and parameter of renal functional magnetic resonance imaging and the pathological changes of the kidney in the previous work, and the application extracts the imaging features of the kidney multi-modal functional magnetic resonance image by using the AI image analysis technology, establishes the relationship between the image and the renal function, constructs the correlation between the imaging and functional genomics, and predicts the model, and demonstrates the correlation between the features of each mode of fMRI and the prediction of senile CKD, thereby providing a reference for the clinical diagnosis and treatment of senile CKD.

[0029] Novel biomarkers of kidney: based on the clinical needs, the biomarkers with the value of predicting the progression of CKD are screened by high-throughput targeted proteomics method combined with machine learning. The biomarkers determined by this method have clinical value, and the prediction accuracy can be improved by including the biomarkers in the prediction model.

[0030] Embodiment 1:

[0031] 1. Selection of research population (inclusion criteria, exclusion criteria) Inclusion criteria: Beijing Hospital nephrology department inpatients or outpatients from 2022 to 2024; age >= 60 years old; meet the diagnosis criteria of CKD 3-4 stage; eGFR 15-60 ml / min / 1.73m 2 ; voluntarily participate in this clinical study, and the patient signs a written informed consent form before the study. The above conditions are met at the same time, that is, the inclusion criteria are met.

[0032] Exclusion criteria: Meet the diagnosis of acute kidney injury; meet the diagnosis of acute kidney injury based on CKD; have a history of kidney transplantation and / or kidney replacement therapy; genetic kidney disease; kidney tumor and tumor resection history; acute and chronic kidney infection; within 2 weeks after acute infection, trauma, surgery; patients with mental illness or those who do not cooperate with the operation; patients with severe internal diseases such as coronary heart disease, emphysema, asthma or those who cannot tolerate breath holding due to other reasons; patients who cannot undergo nuclear magnetic resonance examination, such as claustrophobia, pacemaker or other metal materials; pregnant, lactating or about to be pregnant, lactating during the study; poor compliance (judged by the researchers); and other conditions considered unsuitable for inclusion by the researchers. Any of the above conditions is met, that is, the exclusion criteria are met.

[0033] Since the research of the present application is a prospective observational study, there is no control and randomization, and a third party blind method (blinding the statistical personnel) is used.

[0034] 2. Research process and follow-up plan (as shown in Figure 2 ) Inclusion criteria: hospitalized or outpatient elderly patients with CKD who meet the inclusion and exclusion criteria: ①Diagnosis of CKD meets the 2012 KDIGO standard ②Collect baseline blood and urine samples for molecular biomarker detection ③Synchronously obtain clinical data related to kidney disease and background diseases ④Synchronously perform kidney function magnetic resonance imaging examination ⑤Synchronously perform kidney function examination (glomerular function, tubular function) ⑥Effective follow-up for one year, with ≥2 times of effective serum creatinine measurement within one year ⑦Complete the same items as baseline within the same period of follow-up year ⑧Repeat kidney function magnetic resonance imaging examination for patients who reach the kidney composite endpoint in the follow-up year ⑨Establish complete baseline and follow-up data 3. Study variables 1) Kidney clinical phenotype: Clinical indicators: gender, age, height, weight, BMI, systolic blood pressure, diastolic blood pressure, fasting blood glucose, glycosylated hemoglobin, cholesterol, triglycerides, low-density lipoprotein, high-density lipoprotein.

[0035] Comorbidities [ICD9 / 10 code]: heart failure, coronary heart disease, arrhythmia, myocardial infarction, hypertension, hyperlipidemia, peripheral atherosclerosis, cerebral hemorrhage, cerebral infarction, chronic obstructive pulmonary disease, asthma, hyperuricemia, gout, metabolic acidosis, hyperkalemia.

[0036] Personal history: smoking history, drinking history.

[0037] Concomitant medication: insulin, metformin, sulfonylurea drugs, renin-angiotensin-aldosterone system inhibitors (RAASi) / mineralocorticoid receptor antagonists (MRA), ARB, CCB, diuretics, beta blockers, non-steroidal anti-inflammatory drugs (NSAIDs), calcineurin inhibitors, etc.

[0038] 2) Kidney function phenotype Glomerular function: serum creatinine (calculate eGFR, CKD-EPI formula), serum Cystatin C, urine total protein / creatinine (UPCR), urine albumin / creatinine (UACR), 24-hour urine protein quantification.

[0039] Tubular function: urine NAG / creatinine, urine β2MG, urine α1MG, urine transferrin, urine free light chain κ and λ.

[0040] 3) Kidney molecular phenotype Blood and urine target biomarker screening: Target markers include the following 16: Primary biomarkers: TNFR1, TNFR2, KIM-1, NGAL; Secondary biomarkers: Cystatin C, ADMA, L-FABP, TIMP-2, IGFBP-7, alpha 1M, UMOD, IL-18, MCP-1, YKL-40, PIIINP, EGF.

[0041] Urine sample collection, storage, transportation as follows: All urine sample collection tubes are uniformly distributed, and blank labels are also distributed at the same time; Take the first urine sample of the patient in the morning on an empty stomach. Avoid strenuous exercise such as climbing stairs and drinking a lot of water before collection. Ask the patient to rest for 15 minutes before collection; The collected urine is midstream urine, with a volume of 10-15 mL; Add 100 uL of preservative (boric acid), which can be added before collection; The sample needs to be transported to the laboratory by cold chain after collection.

[0042] Urine sample processing: After the urine sample is left at room temperature for 30 minutes, centrifuge it at 3000 rpm for 10 minutes, then aspirate the supernatant and store it in a 5 ml cryotube, which is stored in a -80°C freezer.

[0043] Blood sample collection, processing and storage: All blood sample collection tubes are uniformly distributed, and blank labels are also distributed at the same time; Take the patient's venous blood in the morning on an empty stomach. Avoid strenuous exercise such as climbing stairs and drinking before collection. Ask the patient to rest for 15 minutes before collection; Collect whole blood 5 mL*2 tubes with EDTA blood collection tubes; After standing at room temperature for 30 minutes, centrifuge at 3000 rpm for 10 minutes at room temperature, then aspirate the supernatant and store it in a 2 ml cryotube; After sample processing, store it in a -80°C freezer; The sample is transported to the laboratory by cold chain.

[0044] Detection method and quantitative system of new biomarker small molecule protein: Detection method: The target small molecule markers of blood and urine samples are quantitatively detected by two methods. ① Utilize high-throughput mass spectrometry targeted proteomics technology to capture large-scale, fast and targeted biomarker proteins, and quantitatively detect these proteins. ② Apply high-sensitivity, wide-linear-range multi-component electrochemical luminescence enzyme-linked detection method to quantitatively detect target biomarker serum proteins.

[0045] Screening and validation of novel biomarkers: The results of targeted analysis were imported into data analysis software Skyline, correct mass spectrometry peaks were screened, and the quantitative analysis results of samples were exported; For the above related molecules, the ROC curve method was used to calculate the optimal threshold value and the corresponding area under the curve of each molecule. The four table method was used to calculate the sensitivity, specificity, positive predictive value and negative predictive value, and the optimal molecular marker was screened out; The logistic regression model of unordered multi-classification was established, the selected molecules were further screened by LASSO penalty algorithm, the prediction model was established by machine learning algorithm, the effect of the model was evaluated by cross-validation method, and the optimal model for predicting the progression of CKD was constructed.

[0046] 4) Kidney imaging phenotype: obtained by kidney functional magnetic resonance imaging examination ① Multi-modal MRI data acquisition: GE 750 3.0T magnetic resonance imaging instrument of Beijing Hospital Department of Radiology was used to collect multi-modal MRI data of all subjects, and the selection of imaging parameters was as follows: BOLD: mGRE sequence was used; IVIM: SS-EPI sequence was used; DKI: SS-EPI sequence was used; PWI: PWI sequence was used; ② Multi-modal MRI data processing: BOLD: The slope of the straight line fitted with the logarithm of the signal intensity of 8 echo T2WI and TE was calculated by matlab7.0 software package, and the R2* value of the corresponding pixel point was obtained, with the unit of 1 / s.

[0047] IVIM, DKI, PWI: The workstation of GE 750 3.0T magnetic resonance instrument was used for post-processing.

[0048] ③ Kidney functional magnetic resonance imaging representation learning and phenotype construction: fMRI multi-modal image data is 2D image data, which is significantly different from the data structure of molecular biomarkers, clinical data, and kidney function detection, so it is necessary to establish a model to convert fMRI image data into low-dimensional representation to participate in the establishment of the prediction model. The specific scheme is: using python programming language combined with pytorch differentiable programming software package, establishing a self-supervised representation (SSR) learning deep learning model to convert fMRI images into low-dimensional image representation, which is used as fMRI image phenotype group. Relying on the correlation between fMRI low-dimensional image representation and the difference of kidney function, and then demonstrating the completeness and clinical feasibility of relying on magnetic resonance imaging to non-invasively predict the progression and prognosis of elderly CKD, and establishing the imaging technology of elderly kidney function decline.

[0049] 5) Evaluation index of follow-up endpoint Determination and definition of kidney composite endpoint: eGFR is calculated by CKD-EPI formula, and the main composite endpoint consists of three parts, which meets one of them.

[0050] Acute kidney function decline [RKFD]: eGFR slope decline ≥5ml / min / 1.73m 2 / year; eGFR decline value ≥40% of baseline value; Entering end-stage renal disease eGFR ≤15ml / min / 1.73m 2 .

[0051] 6) Sample size calculation According to the general principles of risk factor screening design, we estimate that the selected variables of our model are about 25, with 10-20 times of the selected variables as the standard, and then divided by the incidence of eGFR slope decline ≥5ml / min / 1.73m 2 in elderly CKD patients 45% [mainly based on our previous research results to determine] per year, and considering 15% of information missing and other factors, the sample size of this study needs to be between 639-1277 cases. Because the final selected variables of the actual model will be lower than the estimated variables, so the sample size of at least 639 cases can meet the needs of this study.

[0052] 7) Establishment and verification of prediction model The establishment of CKD prediction model can be divided into four parts: database establishment, data preprocessing, feature selection, and prediction model establishment, and the process is shown in Figure 1 .

[0053] The renal clinical phenotype data (including: clinical indicators, comorbidities, personal history, combined medication) of the CKD patients, the renal function phenotype data (glomerular function and tubular function), the renal molecular phenotype data (the concentration of the screened new biomarkers), and the renal imaging phenotype data of the CKD patients are standardized and filled into the database to establish the CKD database, which is used for screening the influencing factors related to CKD and predicting the renal function progression of the patients.

[0054] ①Data in the database are subjected to data preprocessing operations such as data cleaning and data discretization. Data cleaning is a process of reviewing and checking data, including noise data cleaning, missing value processing, etc., discovering and correcting errors in the data, which is conducive to ensuring the consistency and effectiveness of the data. Dirty data, i.e. invalid data and error data for the study, are deleted, and key missing data are filled to obtain a complete and effective data set, which directly affects the effect and reliability of the model. Data discretization is to eliminate the influence of the dimensions of different indicators on the results, and to prepare data for feature extraction.

[0055] ②Machine learning algorithms are used for feature selection (e.g. genetic algorithm, simulated annealing algorithm, etc.), the renal phenotype group variables of the patients are taken as inputs, and the composite renal endpoint is taken as output, so as to select the influencing factors related to CKD for subsequent model prediction; through feature selection, relevant factors are preliminarily screened, irrelevant variable factors are excluded, the input dimension of the prediction model is effectively reduced, the prediction efficiency of the model is improved, and the overfitting of the prediction model is reduced. After feature selection, the influencing factors related to CKD are obtained, but it is not known which (which) factor has a greater influence on CKD and which (which) factor has a smaller influence on CKD, and the relationship between these variables and the glomerular filtration rate of the patients cannot be directly analyzed, therefore, a prediction model needs to be established to predict the glomerular filtration rate of the patients. The methods for establishing the model include: multilayer perceptron (MPL), support vector machine (SVM), random forest (RF), Adaboost algorithm, etc.

[0056] ③Establish a CKD prediction model, the input of the model is n factors related to CKD, and the output of the model is the composite kidney endpoint of the patient; in the training process, the data needs to be divided into three parts: training set, validation set and test set. The training set is used to train the training parameters of the model; the validation set is used to verify the performance of the model, and the model parameters can be adjusted according to the validation result; the test set is the same as the validation set, and can be used to verify the generalization performance of the final trained model, and in general, the ratio of the training set and the validation set is 8:2. In order to verify the performance of the trained model, the following indicators can be used to evaluate: accuracy, sensitivity, specificity, precision, ROC curve, etc. The accuracy is the number of samples correctly classified divided by the total number of samples. Generally speaking, the higher the accuracy, the better the classifier. Sensitivity indicates the proportion of positive examples correctly classified, which measures the recognition ability of the classifier for positive examples. Specificity indicates the proportion of negative examples correctly classified, which measures the recognition ability of the classifier for negative examples. Precision indicates the proportion of examples classified as positive that are actually positive. The ROC curve is a curve with false positive rate and true positive rate as axes. The ROC curve of the clinical + function + biomarker + nuclear magnetic data set in the random forest model is shown in FIG. 8. Figure 3

[0057] ④Output and comparison of prediction model: the present application will construct four-dimensional kidney phenotype group data, including clinical phenotype, molecular phenotype, functional phenotype and imaging phenotype. It is estimated that based on the same kidney composite endpoint, different kidney phenotype groups will be combined in different ways, including clinical + functional phenotype combination, functional + molecular phenotype combination, clinical + molecular + imaging phenotype combination, and clinical + functional + molecular + imaging phenotype combination. The models produced by different combinations may be different in positive predictive value and negative predictive value, which may produce different CKD progression prediction models. Through the same artificial intelligence machine learning method, different phenotype group combination data can be applied to construct CKD prediction models in the same cohort, and the specificity and sensitivity of different models can be verified and compared to determine the most suitable prediction model for clinical application.

[0058] The parts of the present application not described in detail can refer to the prior art or be known to those skilled in the art, and the present embodiment does not limit them, and they will not be described in detail here.

[0059] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative and not limiting. Those skilled in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, which are all within the protection of the present application.​

Claims

1. A method for constructing a predictive model for the progression of chronic kidney disease in the elderly based on renal phenotypic atlas, characterized in that, The method comprises: S1. Basic design: determine the inclusion criteria and exclusion criteria, define the study population; determine the type of study and blind design; S2. Four-dimensional kidney phenotype construction: through the standardization of the whole process of data acquisition-processing-storage, construct four phenotype databases of clinical, molecular, functional and imaging, covering multi-dimensional data from macro clinical information to micro molecular markers; S3. Develop follow-up plan and define endpoints; S4. Sample size calculation: based on the double logic of risk factor screening and outcome incidence, determine the minimum sample size; S5. AI prediction model construction and verification: use classic machine learning algorithms to build models, evaluate model performance through multi-dimensional indicators to ensure prediction accuracy; build models of different phenotype combinations; S6. Model comparison and screening: compare specificity and sensitivity to select the optimal clinical application model from the multiple models constructed. 2.The method of claim 1, wherein the method is characterized by, The four-dimensional kidney phenotype of step S2 comprises: kidney clinical phenotype, kidney molecular phenotype, kidney functional phenotype and kidney imaging phenotype. 3.The method of claim 2, wherein the method is characterized by, The data content of the kidney clinical phenotype comprises: Basic clinical indicators: gender, age, height, weight, BMI, systolic blood pressure, diastolic blood pressure, fasting blood glucose, glycosylated hemoglobin, cholesterol, triglyceride, low-density lipoprotein and high-density lipoprotein; Comorbidities: heart failure, coronary heart disease, arrhythmia, myocardial infarction, hypertension, hyperlipidemia, peripheral atherosclerosis, cerebral hemorrhage, cerebral infarction, chronic obstructive pulmonary disease, asthma, hyperuricemia, gout, metabolic acidosis and hyperkalemia; Personal history: smoking history and drinking history; Concomitant medication: insulin, metformin, sulfonylurea drugs, renin-angiotensin-aldosterone system inhibitors, mineralocorticoid receptor antagonists, ARB, CCB, diuretics, beta blockers, non-steroidal anti-inflammatory drugs and calcineurin inhibitors; The core biomarkers of the kidney molecular phenotype comprise co-targeting 16 biomarkers related to CKD progression; specifically comprising: Primary biomarkers: tumor necrosis factor receptor 1, tumor necrosis factor receptor 2, kidney injury molecule 1 and neutrophil gelatinase-associated lipocalin; Secondary biomarkers: cystatin C, asymmetric dimethylarginine, liver-type fatty acid-binding protein, tissue inhibitor of metalloproteinase 2, insulin-like growth factor binding protein 7, alpha 1-microglobulin, urine albumin, interleukin 18, monocyte chemotactic protein 1, chitinase 3-like protein 1, type III procollagen amino-terminal peptide and epidermal growth factor; The data of the kidney functional phenotype comprises: Glomerular function indicators: serum creatinine, serum Cystatin C, urine total protein / creatinine, urine albumin / creatinine, 24-hour urine protein quantification; Tubular function indicators: urine NAG / creatinine, urine beta 2-microglobulin, urine alpha 1-microglobulin, urine transferrin, urine free light chain kappa and lambda; The kidney imaging phenotype: based on multi-modal MRI technology, construct non-invasive and quantitative kidney imaging data, focusing on the evaluation of pathophysiological states including renal fibrosis and blood perfusion. 4.The method of claim 1, wherein the method is characterized by, The content of the follow-up plan in step S3 includes: the follow-up period is 1 year, and the number of valid serum creatinine measurements within 1 year is required to be ≥2 times to ensure that the eGFR trend can be traced back; all detection items in the baseline period are repeated at the end of the follow-up period; for patients who reach the kidney composite endpoint, repeat the kidney multi-modal MRI examination to evaluate the changes in the image phenotype; The endpoint is the kidney composite endpoint, and eGFR is calculated by CKD-EPI formula. Any one of the following conditions is judged as reaching the endpoint: Acute kidney function decline: eGFR slope decrease ≥ 5 ml / min / 1.73 m 2 per year; The absolute value of eGFR decreases by ≥40% of the baseline value; End stage renal disease: eGFR < 15 ml / min / 1.73m 2 . 5.The method of claim 1, wherein the method is characterized by: The sample size calculation content of step S4 includes: according to the general principle of risk factor screening design, estimating the number of selected variables of the model, taking 10-20 times of the number of selected variables as the standard, and then dividing by the incidence of eGFR slope decline ≥5ml / min / 1.73m 2 in old CKD patients per year, and considering the missing information factor, the value of the sample size is calculated. 6.The method of claim 1, wherein the method is characterized by, The steps of constructing the AI prediction model in step S5 include: S51. Establish a standardized database: standardize the four-dimensional phenotype data of clinical, molecular, functional and imaging, associate with the unique identification of patients, form a structured CKD database, and use it for subsequent modeling; S52. Data preprocessing: clean the data and perform data discretization; S53. Feature selection: use machine learning algorithms, take four-dimensional phenotype variables as input, and take kidney composite endpoint as output, to screen features significantly related to CKD progression and exclude irrelevant variables; S54. Model training: divide the obtained feature data into training set, validation set and test set according to the ratio of 7:1:2, train the classical machine learning algorithm model, optimize the parameters, adjust and evaluate the generalization ability. 7.The method of claim 6, wherein the method is characterized by, The classical machine learning algorithm includes multilayer perceptron, support vector machine, random forest and Adaboost algorithm. 8.The method of claim 1, wherein the method is characterized by, The verification in step S5 is to evaluate the model performance by multi-dimensional indicators to ensure the prediction accuracy; the multi-dimensional indicators specifically include: Accuracy: the number of correctly classified samples / total number of samples, reflecting the overall prediction ability; Sensitivity: true positive sample number / (true positive+false negative), reflecting the recognition ability of patients who have progressed to the endpoint; Specificity: true negative sample number / (true negative+false positive), reflecting the recognition ability of patients who have not progressed; Precision: true positive sample number / (true positive+false positive), reflecting the reliability of positive prediction results; ROC curve: the false positive rate as the horizontal axis and the true positive rate as the vertical axis to draw the curve, the larger the AUC value, the stronger the model discrimination ability. 9.The method of claim 1, wherein the method is characterized by: In step S5, four different phenotype combination models are constructed, including: Model 1: clinical phenotype+functional phenotype, basic combination, covering traditional clinical indicators; Model 2: functional phenotype+molecular phenotype, focusing on function+molecule, strengthening the correlation of micro-mechanism; Model 3: clinical phenotype+molecular phenotype+imaging phenotype, excluding functional phenotype, verifying the synergistic effect of multi-modal; and Model 4: clinical phenotype+functional phenotype+molecular phenotype+imaging phenotype, full-dimensional combination. 10.The method of claim 9, wherein the method is characterized by: The content of step S6 includes: verifying by the same queue data, comparing the verification indicators of the four models, selecting the model with the best performance as the old CKD progression prediction model, and using it for clinical practice.

Images