A method for constructing a chronic kidney disease patient heart failure and all-cause death risk prediction model

Abstract

This invention discloses a method for constructing a predictive model for heart failure and all-cause mortality risk in patients with chronic kidney disease (CKD). It relates to the field of model construction technology, and its key technical points are as follows: This invention uses G3a-G5 stage CKD patients with preserved ejection fraction as the research subjects, establishes a prospective cohort, collects clinical indicators, biomarkers, and quality of life scores, and uses new-onset heart failure combined with all-cause mortality as the composite endpoint. After standardized follow-up and data preprocessing, univariate Cox regression, Lasso, random forest, XGBoost combined with Venn diagrams are used to screen predictive factors. Age, lipoprotein(a), ferritin, GDF15, and EQ-5D quality of life score are identified as independent predictors. A time-dependent nomogram model is constructed, and multi-dimensional efficacy validation is performed. This model is adapted to the specific pathophysiological state of CKD, with accurate predictions and good calibration. With accompanying online interactive tools, it can achieve individualized risk assessment and stratified intervention for 24-36 months, solving the problems of poor applicability and inaccurate prediction in existing models, and has strong clinical applicability.

Description

Technical Field

[0001] This invention relates to the field of model building technology, specifically to a method for constructing a model for predicting the risk of heart failure and all-cause mortality in patients with chronic kidney disease. Background Technology

[0002] Chronic kidney disease (CKD) has become a prevalent chronic non-communicable disease worldwide, characterized by its prolonged course and numerous complications, posing a serious threat to public health. Patients with advanced CKD (stages G3-G5) have a significantly higher risk of cardiovascular death than the general population. Heart failure is its most common and fatal complication, and the leading cause of repeated hospitalizations, a sharp decline in quality of life, and even death. Compared to the general population, the mechanisms of CKD complicated by heart failure are more complex, involving multiple factors such as volume overload, neuroendocrine activation, mineral and bone metabolism disorders, chronic inflammation, and oxidative stress. In particular, CKD plays a dominant role in patients with heart failure with preserved ejection fraction (HFpEF), and the prognosis is generally worse.

[0003] However, current clinical tools for predicting cardiovascular risk in CKD patients, such as the Framingham Risk Score and the ACC / AHA Heart Failure Risk Calculator, are based on the general population or individuals with cardiovascular disease, lacking adaptation to the specific pathophysiological state of CKD. These models primarily rely on traditional risk factors such as age, blood pressure, and blood lipids, failing to fully integrate key dimensions of CKD, including renal function indicators, inflammatory microenvironment, metabolic disorders, and impaired quality of life. Consequently, they often underestimate or overestimate risk in the CKD population, exhibiting insufficient discriminative power and calibration, making it difficult to meet the needs of precise clinical assessment.

[0004] Therefore, developing a dedicated heart failure and all-cause mortality risk prediction model specifically for G3a-G5 stage CKD patients with preserved ejection fraction, integrating routine clinical indicators, novel biomarkers, and quality of life scores, has become an urgent clinical need and has significant medical and social value. Summary of the Invention

[0005] The purpose of this invention is to solve the above-mentioned problems by providing a method for constructing a prediction model for the risk of heart failure and all-cause mortality in patients with chronic kidney disease, so as to solve the problems of insufficient applicability of existing prediction models and limitations in model structure and practicality.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A method for constructing a risk prediction model for heart failure and all-cause mortality in patients with chronic kidney disease, comprising the following steps: Step (1): Including patients with G3a-G5 stage chronic kidney disease with preserved ejection fraction, excluding patients with a history of heart failure, severe cardiovascular disease, and acute inflammatory state, and establishing a prospective study cohort; Step (2): Collecting patients' demographic characteristics, clinical indicators, laboratory test indicators, echocardiographic parameters, and quality of life scale scores; Step (3): Using new-onset heart failure combined with all-cause mortality as the primary composite endpoint, completing a follow-up of up to 36 months and determining the outcome event; Step (4): Performing data missing data analysis. Value imputation, normality test and dataset partitioning, divided into training set and validation set in a 7:3 ratio; Step (5): Using univariate Cox regression, Lasso regression, random forest and XGBoost algorithm combined with Venn diagram cross-analysis to determine candidate predictors; Step (6): Based on multivariate Cox proportional hazards regression, age, lipoprotein(a), ferritin, GDF15 and EQ-5D quality of life score are used as independent predictors to construct a time-dependent nomogram model; Step (7): The model performance is verified by C-index, time-dependent ROC curve, calibration curve, decision curve analysis and Kaplan-Meier survival analysis.

[0007] Furthermore, the ejection fraction mentioned in step (1) is retained as the left ventricular ejection fraction (LVEF) ≥ 45%, and the estimated glomerular filtration rate (eGFR) is calculated using the CKD-EPI formula.

[0008] Furthermore, the laboratory testing indicators mentioned in step (2) include routine blood biochemistry indicators, lipoprotein(a), ferritin, and GDF15; the quality of life scale scores include the European Five-Dimensional Health Scale EQ-5D quality of life score and the Minnesota Heart Failure Quality of Life Questionnaire (MLHFQ) score.

[0009] Furthermore, the follow-up method described in step (3) is a combination of structured telephone interviews, offline clinical visits and medical record verification, with follow-up every 3 months for the first 12 months and every 6 months thereafter.

[0010] Furthermore, the missing value imputation in step (4) adopts the random forest algorithm, and there is no statistical difference in the baseline features of the training set and the validation set after the dataset is divided.

[0011] Furthermore, in step (5), the single-factor Cox regression uses P<0.05 as the threshold to screen potential predictors; Lasso regression uses L1 regularization and 10-fold cross-validation to determine the optimal λ value; and random forest and XGBoost algorithms complete the ranking of feature importance.

[0012] Furthermore, the candidate predictors finally determined by the Venn diagram cross-analysis in step (5) are age, lipoprotein(a), ferritin, GDF15, and EQ-5D quality of life score.

[0013] Furthermore, the multifactor Cox regression described in step (6) uses a two-way stepwise method to screen variables, and the proportional hazards hypothesis is satisfied by the Schoenfeld residual test.

[0014] Furthermore, the nomogram model described in step (6) includes risk probability axes for four time points: 24 months, 27 months, 32 months, and 36 months, and an online dynamic interactive version is also developed.

[0015] Furthermore, the model validation in step (7) uses the Bootstrap method with 1000 resampling cycles to complete internal validation. The validation set is used for internal testing and validation of the model. The model is divided into high-risk and low-risk groups with a risk score of -0.163 as the cutoff value.

[0016] Compared with existing technologies, the beneficial effects of this solution are as follows: This invention targets a specific high-risk clinical population—G3a-G5 stage chronic kidney disease patients with preserved ejection fraction—whose coverage by existing tools is insufficient. It innovatively integrates routine clinical indicators, novel biomarkers, and the EQ-5D quality of life score. Using a multi-algorithm combination of univariate Cox regression, Lasso, random forest, and XGBoost, and cross-validated with Venn diagrams, it identifies five independent predictors: age, lipoprotein(a), ferritin, GDF15, and EQ-5D quality of life score. A time-dependent nomogram model is constructed, fundamentally solving the problems of traditional heart failure and all-cause mortality risk models being based on the general population, not adapting to the specific pathophysiology of CKD, and prone to risk misjudgment. This significantly improves predictive accuracy and population suitability. Tested with Bootstrap internal resampling and internal validation sets, the model shows a high C-index, a continuously increasing AUC, good calibration, and clinical net benefit, enabling accurate long-term dynamic risk assessment over 24-36 months.

[0017] This invention ensures data quality through a prospective cohort approach, standardized data collection, and regulated follow-up. It employs random forests to fill in missing values, a 7:3 dataset partitioning, and a multi-dimensional validation system to ensure model stability and reliability. Furthermore, it develops online dynamic interactive tools and clarifies risk stratification thresholds, transforming complex predictive models into clinically usable, visualized, and operable tools that enable rapid individualized risk assessment without complex calculations. Based on the model's output, it can directly guide stratified intervention and follow-up strategies for low / intermediate / high-risk patients, streamlining the entire chain from model construction to clinical implementation. This significantly improves clinical applicability and scalability, providing an efficient, low-cost, and replicable technical solution for early warning, precise intervention, and prognostic improvement of heart failure in CKD patients. Attached Figure Description

[0018] Figure 1 This is a flowchart of the subject inclusion process in an embodiment of the present invention; it includes the baseline population, exclusion criteria, and subjects to be analyzed;

[0019] Figure 2 This is a Venn diagram of biomarkers identified by the Lasso regression, random forest, and XGBoost models in the embodiments of this invention;

[0020] Figure 3 This is the regression regularization path and the Lasso method for Lambda optimization in the embodiments of the present invention; wherein, Figure A is the coefficient trajectory during the shrinkage process, and Figure B is the cross-validation parameter tuning to obtain the optimal λ value;

[0021] Figure 4 This is the ranking of variable importance in the embodiments of the present invention; wherein, Figure A is the random forest algorithm and Figure B is the XGBoost classifier model;

[0022] Figure 5 This is a nomogram predicting the time-dependent composite endpoint in CKD patients in an embodiment of the present invention; wherein, Figure A shows the nomogram construction including five independent predictors: age, lipoprotein(a), ferritin, GDF15 and EQ-5D score, P<0.05, **P<0.001; Figure B is an interactive online version accessible at: https: / / jiang221.shinyapps.io / dynnomapp;

[0023] Figure 6 This is the time-dependent receiver operating characteristic curve of the training queue at a specified time quantile in this embodiment of the invention;

[0024] Figure 7 It is the area under the ROC curve (AUC) of the training queue as a function of time quantiles in this embodiment of the invention;

[0025] Figure 8 These are the internal verification calibration curves of the training queue at preset time points (24 / 27 / 32 / 36 months) in this embodiment of the invention: where Figure A is for 24 months; Figure B is for 27 months; Figure C is for 32 months; and Figure D is for 36 months.

[0026] Figure 9 This is the time-varying subject operating characteristic curve of the verification queue at a specified time quantile in this embodiment of the invention;

[0027] Figure 10 This is the time-varying subject operating characteristic curve of the verification queue at a specified time quantile in an embodiment of the present invention;

[0028] Figures 11-13These are, respectively, the calibration curve of the verification queue, the decision curve of the training queue, and the decision curve of the verification queue within the pre-set time point line graph in this embodiment of the invention; wherein, Figure A represents 24 months; Figure B represents 27 months; Figure C represents 32 months; and Figure D represents 36 months.

[0029] Figure 14 This is a Kaplan-Meier curve of the overall survival training cohort of CKD patients stratified by risk score in an embodiment of the present invention. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this invention can be combined with each other. The invention will now be described in detail with reference to the embodiments.

[0031] This is a prospective cohort study, with a study period from November 2020 to September 2024, a maximum follow-up period of 36 months, and a median follow-up period of 29 months (IQR 24-35). The research institution is the Department of Nephrology, First Affiliated Hospital of Anhui Medical University. This invention was approved by the Clinical Research Ethics Committee of the First Affiliated Hospital of Anhui Medical University, ethics approval number: PJ2020-11-19; the study design, data collection, and follow-up procedures strictly followed the ethical principles of human research in the Declaration of Helsinki; all subjects signed written informed consent forms before enrollment. The CKD cohort data used in this study are consistent with previously published studies (PMCID: PMC11728072) and fall within the coverage of this ethics approval number, requiring no additional ethical review.

[0032] Example: Method for constructing a predictive model for heart failure and all-cause mortality risk in patients with chronic kidney disease

[0033] I. Screening and Enrollment of Study Subjects

[0034] Inclusion criteria: 1) Patients diagnosed with CKD stage G3a-G5, with estimated glomerular filtration rate (eGFR) calculated according to the Chronic Kidney Disease Epidemiology Collaboration Group (CKD-EPI) formula. The CKD-EPI creatinine equation is: eGFR = 141×min(SCr / κ,1). α ×max(SCr / κ, 1) -1.200 ×0.9938 年龄× Gender coefficient [κ value: 0.7 mg / dL for females, 0.9 mg / dL for males; α value: -0.329 for females, -0.411 for males; gender coefficient: 1 for males, 0.97 for females]; where SCr = standardized serum creatinine; 2) Ejection fraction preserved (LVEF≥45%), confirmed by echocardiography; 3) Age ≥18 years; 4) Able to cooperate in completing baseline clinical tests, laboratory tests, echocardiography and quality of life questionnaires, and able to complete the entire follow-up; 5) Signed written informed consent form.

[0035] Exclusion criteria: 1) Patients diagnosed with heart failure according to Framingham criteria or with a history of symptomatic heart failure (HF); 2) Patients with cerebrovascular events (ischemic / hemorrhagic stroke); 3) Patients diagnosed with atherosclerotic cardiovascular disease (CVD); 4) Patients with moderate to severe valvular disease (echocardiographic confirmation of valvular stenosis / regurgitation ≥ moderate to severe) or a history of valve replacement surgery; 5) Patients with left ventricular dysfunction, LVEF < 45%; 6) Patients with peripheral artery disease; 7) Patients in an acute inflammatory state (such as peritonitis, active autoimmune disease, acute infection, etc.); 8) Patients with incomplete clinical data, unable to cooperate with follow-up, or refusing to sign informed consent.

[0036] The screening and enrollment process includes the following four stages. For details on the subject recruitment process, please refer to [link to relevant documentation]. Figure 1 .

[0037] Phase 1: Initial screening. CKD patients who meet the inclusion criteria are screened through the hospital's electronic medical record system. Medical history, clinical characteristics, and laboratory parameters are cross-referenced to exclude patients who clearly meet the exclusion criteria and to preliminarily determine the candidate group.

[0038] Phase Two: Review, face-to-face clinical assessment of the candidate group, reconfirmation of inclusion / exclusion criteria, collection of basic information, and confirmation of follow-up intentions;

[0039] Phase 3: Enrollment and Testing. For eligible participants who have signed informed consent forms, baseline physical examinations, laboratory tests, echocardiography, and completion of a quality of life questionnaire will be conducted.

[0040] Phase 4: Final enrollment, exclusion of patients who do not meet the criteria during the testing process, numbering of patients who have completed all baseline tests, establishment of a study cohort, and follow-up.

[0041] Enrollment results: 459 patients were initially screened at baseline. Eight patients with LVEF <45% were excluded, leaving 451 patients. Eleven patients were lost to follow-up during the follow-up period, and a total of 440 subjects were finally included for subsequent analysis.

[0042] II. Comprehensive Baseline Data Collection

[0043] Systematic baseline data were collected from the 440 patients who were ultimately enrolled. The data were divided into four categories: demographic and clinical characteristics, laboratory tests, echocardiography, and quality of life scale assessment. All testing procedures were performed by professionals, and the testing equipment and reagents were standardized.

[0044] Demographic and clinical characteristics collection includes demographic information, physical examination indicators, and comorbidities and medication history; demographic information includes age, sex, smoking status (divided into former smokers / never smokers, former smokers include current smokers and past smokers), alcohol consumption status (divided into former alcoholics / never alcoholics, alcoholism is defined as long-term excessive drinking), and peritoneal dialysis status (whether receiving peritoneal dialysis, dialysis duration); physical examination indicators include sitting systolic blood pressure (SBP) and diastolic blood pressure (DBP), measured using a standard electronic blood pressure monitor, consecutively... Three measurements were taken consecutively, and the average value was recorded. Heart rate was measured using an electrocardiograph or a fingertip heart rate monitor. Height and weight were measured, and the body mass index (BMI = weight / height², kg / m²) was calculated. Waist and hip circumference were measured, and the waist-to-hip ratio (WHR) was calculated. Comorbidities and medication history were recorded, including whether hypertension or diabetes was present. Hypertension was defined as an average SBP / DBP ≥ 140 / 90 mmHg or regular use of antihypertensive medication. Diabetes diagnosis was based on the patient's self-reported medical history and the use of hypoglycemic medications / blood glucose test results. The patient's use of antihypertensive, hypoglycemic, and lipid-lowering medications was recorded.

[0045] All subjects fasted overnight (≥8 hours) after admission. The following morning, 10 mL of peripheral blood was drawn from the antecubital vein and placed in a blood collection tube without anticoagulant. At the same time, 5 mL of fresh urine was collected. After the blood samples were allowed to stand for 30 minutes, they were centrifuged at 3000 r / min for 10 minutes to separate the serum. 7 mL of serum was sent to the hospital's clinical laboratory for routine testing, and the remaining 3 mL of serum was placed in a cryovial and transported to the Anhui Medical University Research Center via dry ice. It was stored at -80°C for the detection of specific biomarkers.

[0046] The tests were performed by the hospital's clinical laboratory using a fully automated biochemical analyzer. The test indicators included: (1) Complete blood cell count: red blood cell count (RBC), hemoglobin (HGB), white blood cell count, platelet count (PLT), neutrophils (NEUT), lymphocytes, and eosinophils (EO); (2) Kidney function indicators: eGFR (CKD-EPI formula), blood urea nitrogen (BUN), serum creatinine (CRE), and uric acid (UA); (3) Liver function indicators: total bilirubin (TBIL) and indirect bilirubin (IBIL); (4) Lipid indicators: lipoprotein (a) [Lp(a)]; (5) Electrolytes and minerals: serum calcium, serum phosphorus (P), and serum iron (Fe); (6) Iron metabolism indicators: ferritin (FER); (7) Inflammation and enzyme indicators: alkaline phosphatase, C-reactive protein (CRP), and creatine kinase MB isoenzyme (CK-MB). All test results were extracted from the values ​​recorded in the electronic medical record system, and the reference range of each indicator was also recorded.

[0047] The assays were performed by the research center using enzyme-linked immunosorbent assay (ELISA). The kits were from Elabscience Biotech (Wuhan, China). The procedures were strictly followed according to the kit instructions. The assay indicators included: growth differentiation factor 15 (GDF15), full-length intact fibroblast growth factor 23 (iFGF23), soluble tumor growth inhibitory factor 2 (sST2), galactohexin-3 (GAL3), and α-klotho. All assays were performed in duplicate, and the average value was used as the final result. The coefficient of variation within and between batches was controlled to be <10%.

[0048] (III) Echocardiography

[0049] 1. Testing equipment and personnel

[0050] This invention uses the GE Vivid E95 cardiac ultrasound system, which is operated independently by two physicians with more than 5 years of experience in cardiac ultrasound examination. The test results are the average of the two results, and if there is a disagreement, a third senior physician will make the final decision.

[0051] 2. Testing Standards and Content

[0052] A comprehensive transthoracic echocardiogram was performed in strict accordance with the guidelines of the American Society of Echocardiography (ASE), including: (1) Two-dimensional imaging: acquiring the long axis, two-chamber, and four-chamber views of the heart, and the short axis views of the left ventricle; (2) Left ventricular ejection fraction (LVEF): measured using the modified Simpson dual-plane technique to ensure LVEF ≥ 45%; (3) Left ventricular mass index (LVMi): the left ventricular mass was measured by M-mode ultrasound, and LVMi was calculated by combining body surface area (BSA) correction; (4) Diastolic function parameters: peak diastolic velocity of the mitral valve in early diastole (E wave) and peak diastolic velocity of the mitral valve annulus in early diastole (e' wave), and the E / e' ratio was calculated; (5) Speckle tracking echocardiography (STE): longitudinal strain data of the long axis, two-chamber, and four-chamber views of the heart were acquired, and global longitudinal strain (GLS) was calculated. In this study, the negative sign of GLS was omitted in the report; all test data were entered into the echocardiography workstation in real time and archived by a designated person.

[0053] (iv) Quality of life scale assessment

[0054] After baseline patient testing was completed, trained research nurses provided one-on-one guidance to patients on completing questionnaires, ensuring that patients understood the meaning of the questions and answered them independently. For those unable to complete the questionnaires independently, nurses assisted by providing oral answers. All questionnaires were collected and reviewed on-site to exclude incomplete or incorrectly completed questionnaires. The assessment tools included two validated scales:

[0055] The European Five-Dimensional Health Scale (EQ-5D) (1) Assessment dimensions: mobility, self-care, daily activities, pain / discomfort, anxiety / depression, a total of 5 dimensions; (2) Scoring criteria: each dimension is divided into 5 levels, namely "no problem, minor problem, moderate problem, severe problem, extreme problem / inability to perform"; (3) Result calculation: generate health utility value, ranging from 0 (representing death) to 1 (representing perfect health), and record the visual analog scale (VAS) score.

[0056] Minnesota Heart Failure Quality of Life Questionnaire (MLHFQ) (1) Assessment content: 21 items, divided into physical health dimension (8 items) and emotional health dimension (5 items); (2) Scoring criteria: 0-5 Likert scale, 0 points is "no impact", and 5 points is "serious impact"; (3) Result calculation: Total score range 0-105 points, the higher the score, the worse the quality of life; physical dimension 0-40 points, emotional dimension 0-25 points.

[0057] III. Follow-up and outcome determination

[0058] Follow-up was conducted every 3 months for the first 12 months, and every 6 months thereafter, until the 36-month follow-up ended or an outcome event occurred; This invention adopts a combination of structured telephone interviews, offline clinical visits, and medical record verification; (1) Telephone interviews: Inquire about the patient's physical condition, whether there are HF-related symptoms (such as dyspnea, fatigue, edema, etc.), and hospitalization status; (2) Offline visits: At the follow-up time point, a physical examination, laboratory indicator re-examination, and echocardiogram examination were completed at the hospital; (3) Medical record verification: The patient's hospitalization, treatment, and death records were verified through the hospital's electronic medical record system and medical insurance system; For patients with incomplete questionnaire or interview data, supplementary telephone interviews were conducted with their families; Patients who could not obtain complete data were judged as lost to follow-up and excluded from the study cohort.

[0059] This invention uses new-onset heart failure (HF) plus all-cause mortality as the primary outcome predicted by the model. A consensus panel of three senior cardiologists independently determined the outcome based on the Framingham Heart Failure Diagnostic Criteria, combining all available data including the patient's clinical symptoms, signs, laboratory tests (such as NT-proBNP), and echocardiographic results. A diagnosis was confirmed only if all three physicians agreed. All deaths, regardless of cause, were included in the statistics. The time and cause of death were verified through hospital death records, family confirmation, and household registration cancellation certificates, and were categorized as cardiovascular death and non-cardiovascular death. After follow-up of 440 participants, 104 (23.6%) reached the primary composite endpoint, including 62 (14.1%) new-onset HF and 42 (9.5%) all-cause mortality (12 cardiovascular deaths and 30 non-cardiovascular deaths).

[0060] IV. Data Preprocessing

[0061] All data was created using Excel 2019, with data entry and verification performed by two people to ensure accuracy; data preprocessing was performed using R version 4.2.3 and Python version 3.11.4.

[0062] Missing values ​​were screened for all covariates. In this study, the missing value rate of all covariates was <5%, and there were no large number of missing variables. The Random Forest (RF) method in the Python Scikit-learn library (version 1.1.3) was used to impute missing values. This method is superior to traditional mean imputation and multiple imputation, and can better handle complex nonlinear data structures and maintain the intrinsic relationship between variables. The data was standardized before imputation to ensure the reliability of the imputation results.

[0063] The Shapiro-Wilk test was used to assess the normality of all measurement data. P>0.05 was considered to be in accordance with normality, and P<0.05 was considered to be non-normal. (1) Normally distributed measurement data: expressed as mean ± standard deviation (x±s), and the independent samples t test was used for comparison between groups. (2) Non-normally distributed measurement data: expressed as median (interquartile range) [M (IQR)], and the Mann-Whitney U test was used for comparison between groups. (3) Count data: expressed as frequency (percentage) [n (%)], and the chi-square test (χ²) was used for comparison between groups. The study subjects were divided into the event group (n=104) and the non-event group (n=336) according to whether the major composite endpoint occurred, and the baseline characteristics between groups were compared.

[0064] This invention uses a random number method to divide a complete dataset of 440 patients into a training set and a validation set in a 7:3 ratio: the training set (n=308) is used for further screening of predictive factors, construction of the nomogram model, and internal validation; the validation set (n=132) is used for external independent validation of the model. Baseline clinical characteristics, laboratory indicators, echocardiographic parameters, quality of life scale scores, and incidence of the primary composite endpoint are compared between the training and validation sets to ensure that there are no statistically significant differences between the two groups (P>0.05), thus ensuring the homogeneity of the dataset.

[0065] V. Predictor Selection

[0066] This invention employs a multi-stage analysis framework of "single-factor Cox regression + multiple machine learning algorithms + Venn diagram cross-analysis" to screen out predictive factors that have a stable and significant association with the main composite endpoint, ensuring the predictive value and reliability of the factors. All analyses are performed on the training set.

[0067] Phase 1: Preliminary screening using univariate Cox proportional hazards regression

[0068] All baseline variables were used as independent variables, and the primary composite endpoint (occurrence time + whether it occurred) was used as the dependent variable. A univariate Cox proportional hazards regression analysis was performed, as shown in Table 1 below.

[0069] Table 1. Univariate Cox proportional hazards regression analysis of the composite endpoint

[0070]

[0071] BMI (Body Mass Index); RBC (Red Blood Cell Count); HGB (Hemoglobin); WBC (White Blood Cell Count); PLT (Platelet-Rich Plasma); NEUT (Neutral Cells); lymphocytes; EO (Eosinophils); eGFR (Estimated Glomerular Filtration Rate); BUN (Body Urea Nitrogen); CRE (Cretin); UA (Uric Acid); TBIL (Total Bilirubin); IBIL (Indirect Bilirubin); Lp(a) (Lipoprotein(a)); Ca (Calcium); P (Phosphorus); Fe (Iron); FER (Ferritin); ALP (Alkaline Phosphatase); CRP (C-Reactive Protein); GDF15 (Growth Differentiation Factor 15); FGF (Ferritin Genetic Differentiation Factor). 23, fibroblast growth factor 23; sST2, soluble tumor growth inhibitory factor 2; GAL3, galactolectin-3; NT-proBNP, N-terminal pro-brain natriuretic peptide; CK-MB, creatine kinase-MB isoenzyme; LVEF, left ventricular ejection fraction; E / e' ratio, the ratio of early diastolic inflow velocity (E wave, m / s) to early diastolic velocity (e', m / s) of the mitral valve annulus; LVMi, left ventricular mass index; GLS, global longitudinal strain; MLHFQ, Minnesota Heart Failure Quality of Life Questionnaire score; EQ-5D, European Five-Dimensional Health Scale score.

[0072] Using P<0.05 as the screening threshold, potential predictors that were significantly associated with the primary composite endpoint were screened out. A total of 8 potential predictors were screened out, namely: age, total bilirubin (TBIL), indirect bilirubin (IBIL), lipoprotein (a) [Lp (a)], ferritin (FER), growth differentiation factor 15 (GDF15), Minnesota Heart Failure Quality of Life Questionnaire (MLHFQ) score, and European Five-Dimensional Health Scale (EQ-5D) score.

[0073] Phase Two: Feature Optimization and Selection for Machine Learning Algorithms

[0074] This invention employs three machine learning algorithms for dimensionality reduction, feature extraction, and importance ranking of potential predictors. The hyperparameters of all models are optimized using 10-fold cross-validation and grid search to ensure optimal model performance. This invention uses L1 regularization, aims to minimize the Akaike Information Content Criterion (AIC), sets 10-fold cross-validation, and iterates 1000 times. A coefficient path graph is plotted, as shown below. Figure 3 As shown in Figure A, observe the trend of variable coefficients changing with the regularization parameter λ; determine the optimal λ value (λ=0.098) through cross-validation error curves, as shown in Figure A. Figure 3 As shown in B, the model error is the smallest at this point; finally, four key factors were selected: age, hemoglobin (HGB), GDF15, and E / e' ratio.

[0075] This invention uses Friedman mean squared error (MSE) as the node splitting criterion, sets the number of trees (n_estimators) to 100, the maximum depth (max_depth) to None, and the minimum purity gain (min_impurity_decrease) to 0.0; calculates the feature importance score of each variable, and sorts them from highest to lowest score; extracts the top 10 predictors, in the following order: GDF15, ferritin, E / e' ratio, age, Lp(a), uric acid (UA), CK-MB, α-klotho, RBC, and N-terminal pro-B-type natriuretic peptide (NT-proBNP), etc. Figure 4 As shown in Figure A. The model parameters of this invention are: objective function is squared error (reg), learning rate = 0.1, maximum tree depth = 8, minimum child weight = 6, L2 regularization term (lambda) = 1; the objective is to maximize the area under the receiver operating characteristic curve (AUROC); the feature importance scores of each variable are calculated and sorted from high to low; the top 10 predictors are extracted, in the following order: diabetes, α-klotho, GDF15, EQ-5D score, FGF23, ferritin, BMI, age, LVEF, PLT, etc. Figure 4 As shown in B.

[0076] Phase 3: Venn diagram cross-analysis to identify candidate predictors

[0077] Analysis tools: R software "VennDiagram" package; This invention performs Venn diagram cross-analysis on the eight latent factors of univariate Cox regression and the screening results of three machine learning algorithms: Lasso regression, random forest, and XGBoost, to screen out the intersection variables that are identified in different analytical frameworks, as shown in the appendix. Figure 2 As shown, five candidate predictors were identified: age, GDF15, ferritin, EQ-5D score, and Lp(a). These five factors all showed stable predictive value in the multi-method analysis and were used as input variables for subsequent multivariate analysis.

[0078] VI. Multivariate Cox proportional hazards regression analysis: Identifying independent predictors

[0079] This invention uses five candidate predictors identified by Venn diagram analysis as independent variables and the primary composite endpoint as the dependent variable. Multivariate Cox proportional hazards regression analysis is performed on the training set, employing a two-way stepwise method to screen variables, gradually including / removing variables until the model reaches its optimum. This invention performs Cox proportional hazards hypothesis testing, verifying that the model satisfies the proportional hazards hypothesis (P>0.05) through the Schoenfeld residual test, ensuring model applicability. The regression coefficients (Estimate), standard errors (SE), Z-scores, P-scores, hazard ratios (HR), and 95% confidence intervals (95% CI) for each factor are calculated. Using P<0.05 as the criterion, predictors significantly and independently associated with the primary composite endpoint are identified. All five candidate predictors were confirmed as independent predictors of the primary composite endpoint; specific results are shown in Table 2.

[0080] Table 2 Results of Multivariate Cox Proportional Hazard Regression Analysis

[0081]

[0082] GDF15, growth differentiation factor 15; FER, ferritin; EQ-5D, European Five-Dimensional Health Scale score; Lp(a), lipoprotein(a).

[0083] As shown in Table 1 above, for age (HR=1.030), the risk of new-onset heart failure and all-cause mortality increases by 3.0% for every year of age increase in CKD patients; for GDF15 (HR=1.076), the risk increases by 7.6% for every 100 pg / mL increase in serum concentration; for lipoprotein(a) (HR=1.011), the risk increases by 1.1% for every 10 mg / L increase in concentration; for ferritin (HR=1.009), the risk increases by 0.9% for every 10 μg / L increase in serum concentration; and for EQ-5D score (HR=0.099), the risk decreases by 9.0% for every 0.1 unit increase in score.

[0084] VII. Construction of the Nomogram Prediction Model

[0085] Based on five independent predictors identified through multivariate Cox regression analysis, a time-dependent nomogram prediction model was constructed to achieve individualized and visualized prediction of the risk of new-onset heart failure and all-cause mortality in CKD patients. This invention uses the R software "rms" package (version 6.2.0) for nomogram construction and the "dynnom" package to develop an online dynamic version. Score weights are assigned according to the regression coefficients of each independent predictor; the larger the regression coefficient, the higher the score weight and the stronger the predictive value. The scores of each factor are summed to obtain the total score, which is positively correlated with the probability of the patient experiencing the outcome event, as shown in the appendix. Figure 5As shown in Figure A, the constructed nomogram has a vertical hierarchical structure, from top to bottom as follows: Rating axis: scale 0-100, representing the individual scores and total scores for each factor; Predictor factor axis: representing 5 independent predictors, each factor labeled with its corresponding value range, and the scoring scale intervals are set according to the clinical significance and regression coefficients of the factors; Linear predictive value axis: linear predictive values ​​converted from the total score, scale -2-2.5; Time probability axis: representing four time points at 24 months, 27 months, 32 months, and 36 months, each time point labeled with the probability of occurrence of the corresponding outcome event, scale 0-0.95.

[0086] The specific usage of the nomogram is as follows: Locate the position of the actual measured value / score of each independent predictor factor on the corresponding factor axis; draw a perpendicular line from this position to the left scoring axis and read the individual score of that factor; add the individual scores of the five factors to obtain the total score; draw perpendicular lines from the total score position to the downward linear predictive value axis and each time probability axis, and read the corresponding linear predictive value and the risk probability of new-onset heart failure + all-cause mortality at 24 / 27 / 32 / 36 months. To improve clinical usability, an online dynamic interactive version of the nomogram has been developed, as shown in the attached figure. Figure 5 As shown in B, the access address is: https: / / jiang221.shinyapps.io / dynnomapp; clinicians can directly enter the patient's age, GDF15, ferritin, lipoprotein(a), and EQ-5D score on the webpage, and the system will automatically calculate the individual scores, total score, and risk probability at each time point. There is no need to manually look up the charts, making the operation simple and fast.

[0087] VIII. Comprehensive Validation of Model Performance

[0088] Internal validation of this invention using the Bootstrap method (1000 resampling) yielded a concordance index (C-index) of 0.772 (95% confidence interval 0.725–0.819), indicating that the model has significant predictive power for composite outcomes. Notably, with increasing patient age, GDF15, ferritin, and lipoprotein(a) levels increase, while EQ-5D scores decrease, and cumulative nomogram scores increase accordingly, leading to a higher predictive risk for composite outcomes.

[0089] To assess the impact of time on the prediction results, we conducted time-dependent ROC analyses at pre-defined time points of 24, 27, 32, and 36 months, as shown in the attached figure. Figure 6 As shown. Due to insufficient event count, earlier time points (e.g., 12 months) were not included in the analysis; otherwise, unreliable estimates with excessively wide confidence intervals would have been produced. The dynamic AUC(t) curve is attached. Figure 7As shown, the odds ratio showed a continuous upward trend, from 0.779 (95% CI 0.696–0.863) at 24 months, to 0.816 (95% CI 0.743–0.888) at 27 months, then to 0.836 (95% CI 0.768–0.903) at 32 months, and finally to 0.850 (95% CI 0.771–0.929) at 36 months. (See attached image.) Figure 8 The calibration curves shown confirm a high degree of consistency between the predicted and observed risks of composite events in patients with chronic kidney disease (CKD) at four time points. Meanwhile, the decision curve analysis (DCA) is attached. Figure 12 As shown, the clinical utility of the model becomes more robust as the threshold probability range expands over time, validating its robustness in predicting composite outcomes in the CKD population. This invention uses the optimal risk score cutoff value of -0.163 to divide patients into high-risk and low-risk groups. Log-rank test results show a statistically significant difference in prognosis between the two groups (P<0.001), as shown in the attached figure. Figure 14 As shown in the attached diagram. In the validation set, as shown in the diagram... Figure 9 The ROC curve shown and the attached figure Figure 10 The trend of AUC values ​​shown is consistent with that of the training cohort: the AUC values ​​exhibit a steady upward trend, increasing from 0.648 (95% CI 0.470–0.825) at 24 months to 0.791 (95% CI 0.658–0.924) at 27 months, further rising to 0.884 (95% CI 0.800–0.969) at 32 months, and finally reaching 0.901 (95% CI 0.821–0.982) at 36 months. Similarly, as shown in the attached... Figure 11 The calibration curves shown demonstrate that the predicted and observed values ​​generally exhibit the best agreement. (See attached...) Figure 13 The DCA results shown indicate that the clinical utility of the model becomes more robust as the threshold probability range expands over time, thus validating its robustness in predicting composite outcomes in the chronic kidney disease (CKD) population.

[0090] IX. Clinical Application of the Model

[0091] Based on the risk stratification results of the nomogram model, individualized clinical intervention and follow-up strategies are developed for CKD patients at different risk levels, achieving precision medicine and improving patient prognosis.

[0092] This model uses the optimal risk score cutoff value of -0.163 to divide patients into high-risk and low-risk groups, and the prognosis of the two groups has a statistically significant difference. The nomogram score corresponding to the model risk score of -0.163 is 99 points. Therefore, the nomogram scores obtained by patients can be divided into high-risk (≥99 points) and low-risk (<99 points) groups with a cutoff value of 99 points. In clinical intervention, it can be further subdivided into three levels of management: low-risk group (total score <40 points), medium-risk group (40 points ≤ total score <99 points), and high-risk group (total score ≥99 points), corresponding to different follow-up frequencies and intervention plans. Among them, the low-risk group (total score < 40 points): maintain standard CKD routine care, follow the chronic kidney disease diagnosis and treatment guidelines for basic treatment, including blood pressure, blood sugar, blood lipid control, nutritional support, and optimized dialysis plan for peritoneal dialysis patients; follow up once every 6 months, complete routine physical examination and laboratory tests, complete one echocardiogram examination and EQ-5D score once a year, and promptly detect and intervene in the decline in quality of life. The medium-risk group (40 points ≤ total score < 99 points): on the basis of standard care, implement enhanced lifestyle and metabolic intervention; (1) Dietary intervention: low salt, low fat, low phosphorus diet, reduce lipid intake, and reduce Lp(a) level; (2) Lifestyle intervention: quit smoking and limit alcohol, appropriate exercise, relieve anxiety / depression, and improve EQ-5D score; (3) Drug intervention: optimize antihypertensive and hypoglycemic drugs, add lipid-lowering drugs when necessary, and control inflammation level; follow up once every 3 months, complete routine examinations, complete one echocardiogram examination and EQ-5D score once every 6 months, and closely monitor changes in indicators. High-risk group (total score ≥99 points): On the basis of enhanced intervention, cardio-renal protective targeted therapy is implemented; (1) Iron metabolism intervention: individualized iron therapy is given to patients with functional iron deficiency (transferrin saturation <20%, inflammatory ferritinemia 30-100μg / L) to avoid iron overload; (2) Lipid intervention: RNA targeted therapy is used for patients with good LDL-C control but persistently elevated Lp(a) to reduce the risk of cardiovascular residue; (3) Myocardial protection: avoid the use of nephrotoxic and myocardial toxic drugs, and add myocardial protective drugs when necessary to reduce GDF15 level; follow up once every 1-2 months, complete a comprehensive physical examination, laboratory test and echocardiogram examination, complete EQ-5D score once a month, and adjust the treatment plan in a timely manner.

[0093] The above specific embodiments are merely explanations of the present invention and are not intended to limit the present invention. After reading this specification, those skilled in the art can make modifications to these embodiments without contributing any inventive step, but as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A method for constructing a predictive model for heart failure and all-cause mortality risk in patients with chronic kidney disease, characterized by: Includes the following steps: Step (1): Include patients with G3a-G5 stage chronic kidney disease with preserved ejection fraction, exclude patients with a history of heart failure, severe cardiovascular disease and acute inflammatory state, and establish a prospective study cohort; Step (2): Collect patient demographic characteristics, clinical indicators, laboratory test indicators, echocardiographic parameters, and quality of life scale scores; Step (3): With new-onset heart failure combined with all-cause mortality as the primary composite endpoint, complete a follow-up period of up to 36 months and determine the outcome event; Step (4): Impute missing values, perform normality tests, and split the dataset into training and validation sets in a 7:3 ratio; Step (5): Use univariate Cox regression, Lasso regression, random forest, and XGBoost algorithms in conjunction with Venn diagram cross-analysis to determine candidate predictors; Step (6): Based on multivariate Cox proportional hazards regression, a time-dependent nomogram model is constructed with age, lipoprotein(a), ferritin, GDF15, and EQ-5D quality of life score as independent predictors. Step (7): The model performance was validated through C-index, time-dependent ROC curve, calibration curve, decision curve analysis and Kaplan-Meier survival analysis.

2. The construction method as described in claim 1, characterized in that: The ejection fraction mentioned in step (1) is retained as the left ventricular ejection fraction (LVEF) ≥ 45%, and the glomerular filtration rate (eGFR) is estimated using the CKD-EPI formula.

3. The construction method as described in claim 1, characterized in that: The laboratory test indicators mentioned in step (2) include routine blood biochemistry indicators, lipoprotein(a), ferritin, and GDF15; the quality of life scale scores include the European Five-Dimensional Health Scale EQ-5D quality of life score and the Minnesota Heart Failure Quality of Life Questionnaire (MLHFQ) score.

4. The construction method as described in claim 1, characterized in that: The follow-up method described in step (3) is a combination of structured telephone interviews, offline clinical visits and medical record verification. Follow-up is conducted every 3 months for the first 12 months and every 6 months thereafter.

5. The construction method as described in claim 1, characterized in that: The missing value imputation in step (4) uses the random forest algorithm. After the dataset is split, there is no statistical difference in the baseline features of the training set and the validation set.

6. The construction method as described in claim 1, characterized in that: In step (5), the single-factor Cox regression uses P<0.05 as the threshold to screen potential predictors; Lasso regression uses L1 regularization and 10-fold cross-validation to determine the optimal λ value; and random forest and XGBoost algorithms complete the ranking of feature importance.

7. The construction method as described in claim 1, characterized in that: The candidate predictors finally determined by the Venn diagram cross-analysis in step (5) are age, lipoprotein(a), ferritin, GDF15, and EQ-5D quality of life score.

8. The construction method as described in claim 1, characterized in that: The multifactor Cox regression described in step (6) uses a two-way stepwise method to screen variables, and the proportional hazards hypothesis is satisfied by the Schoenfeld residual test.

9. The construction method as described in claim 1, characterized in that: The nomogram model described in step (6) includes risk probability axes for four time points: 24 months, 27 months, 32 months, and 36 months, and an online dynamic interactive version is also developed.

10. The construction method as described in claim 1, characterized in that: The model validation described in step (7) uses the Bootstrap method with 1000 resampling cycles to complete internal validation. The validation set is used for internal testing and validation of the model. The model is divided into high-risk and low-risk groups with a risk score of -0.163 as the cutoff value.

Images