CKD-M2 study: 2-year mortality prediction tool for advanced kidney disease Authors: Dung N T Tran, Yves Dimitrov, Francois Chantrel, Peggy Perrin, Zead Tubail, Denis Fouque, Michel Ducher, Jean-Pierre Fauvel Published: 2025 ABSTRACT AbstractBackgroundA few models have been developed in recent years to predict all-cause mortality in patients with chronic kidney disease (CKD). However, many have been developed using inappropriate methods and have not been externally validated. This study aims to improve our previously validated tool for predicting 2-year all-cause mortality in stage 4–5 CKD patients by enlarging the training dataset, thereby enhancing its robustness, which is further supported by a second external validation.MethodThe Bayesian network-based 2-year all-cause mortality prediction tool was trained on a comprehensive national dataset, which was further enriched by incorporating data from a previous external validation. Internal performance was assessed using 10-fold cross-validation, while external validity was confirmed through a second validation on the CERRENE cohort. The discriminatory ability of the prediction tool was evaluated both internally and externally using accuracy, c-statistic, sensitivity, and specificity. The calibration of the external validation was visualized with a calibration curve.ResultsThe prediction tool was developed using a training dataset of 1,061 patients (median age 72.3 years; 2-year mortality rate 21.2%) and externally validated using data from 409 patients (median age 77.5 years; 2-year mortality rate 17.6%) with CKD stage 4 or 5. The tool demonstrated satisfactory performance in both internal and external validation (accuracy: 77.2% and 77.8%; AUC-ROC: 0.76 and 0.74; sensitivity: 47.1% and 54.2%; specificity: 85.3% and 82.3%, respectively). The calibration curve demonstrated acceptable agreement between predicted and observed outcomes, and Brier score = 0.132.ConclusionThe updated prediction tool demonstrated satisfactory performance in both internal and external validation processes. Before it can be used in clinical practice, it must undergo national and international external validations, which are currently in progress.Graphical abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s12882-025-04715-x.Keywords: Chronic kidney disease, Mortality prediction, External validation, Machine learning Background The number of deaths associated with chronic kidney disease (CKD) increased by 95% between 2000 and 2021, making CKD rise from the world’s 19th leading cause of death to the 9th in this period [1]. CKD is one of the noncommunicable diseases that have shown an increase in mortality rates since 1990 [2]. The mortality rate increases as the estimated glomerular filtration rate (eGFR) decreases [3]. Age is strongly associated with mortality rates in patients with CKD. For example, the mortality rate is 10.7 per 1,000 person-years for individuals aged 21–35 years, and 67.7 per 1,000 person-years for those aged 66–75 years [4]. In the context of an aging population and the rise of individualized medicine aimed at guiding optimal treatment choices and preparing patients for potential kidney replacement therapy, the accurate assessment of mortality risk in CKD patients has emerged as an unmet need in their clinical management. Externally validated models that predict all-cause mortality specifically developed for the CKD population, are recommended for use to guide nephrologists in discussing care goals [5]. The most widely used prediction tool for patients with CKD aims to predict the risk of end-stage kidney disease but not mortality [6]. Several models have been developed to predict all-cause mortality in patients with CKD. While many of these models were developed using inappropriate methods and have generally been poorly reported [7, 8], recent advancements in health data, medical research, and machine learning have led to the development of more robust models [9–14]. Nevertheless, to the best of our knowledge, no existing tool that specifically predicts 2-year all-cause mortality for patients with stage 4–5 CKD has undergone external validation. Recently, our research group proposed an all-cause mortality prediction tool for patients with stage 4–5 CKD at 2 years, using a Bayesian network and synthetic data [15]. The predictors included in the proposed all-cause mortality prediction tool are accessible in a timely manner and easy to collect in clinical practice [15]. Although the proposed prediction tool showed good discrimination, with an area under the receiver operating characteristic curve (AUC-ROC) of 0.81 in internal validation and 0.72 in external validation, it has modest accuracy [15, 16]. The aim of this study is to improve the Bayesian network-based 2-year all-cause mortality prediction tool by augmenting the training dataset, and then to validate it externally using a second independent database of patients with stage 4–5 CKD. Methods Bayesian network-based all-cause mortality prediction tool for patients with stage 4 and 5 CKD at 2 years During the development phase of this Bayesian network-based 2-year all-cause mortality prediction tool, 26 variables were initially included in the analysis. The contribution of each variable to 2-year all-cause mortality was evaluated, and variables were successively added to the Bayesian network model according to their degree of contribution to the outcome. After each addition, the model’s AUC-ROC value was assessed. To ensure the practical utility of the prediction tool in clinical settings, only a limited number of variables, those that are most important for the outcome, and still yield a satisfactory AUC-ROC value, were included in the model. Consequently, seven variables were included in the model, ranked in descending order of their contribution to the outcome as follow: age, use of erythropoiesis-stimulating agents (ESA), cardiovascular (CV) history (stroke and transient ischemic attack, hospitalization for heart failure, acute coronary syndrome), smoking status (actual, former, never), levels of 25-OH vitamin D, parathyroid hormone (PTH), and ferritin [15]. The Bayesian network model was chosen because it showed better performance in 10-fold cross-validation compared to three other models: deep learning, logistic regression, and random forest. The development process has been described in detail in our previously published model development study [15]. The 7 predictors in the Bayesian network are represented as 7 nodes in the model, connected to mortality through a directed acyclic graph. The continuous predictors were categorized into binary or polynomial variables. For each predictor, a conditional probability was quantified for each category. The probability of mortality, derived from the Bayesian network, is the joint probability calculated from different combinations of the possible values of the 7 predictors, depending on patient’s specific data. The Bayesian network model of the proposed 2-year all-cause prediction tool is described in Supplemental Fig. 1. The combination of age > 84.9 years, PTH level > 150 ng/L, 25-OH vitamin D level < 25 nmol/L, ferritin level 100–299 µg/L, former smoker status, having a cardiovascular history, and treatment with ESA results in the highest probability of mortality [15]. Training dataset The training dataset consisted of two subsets: (i) data from the Photo-Graphe 3 study, which was used to develop the initial prediction tool [15–17]; and (ii) comprehensive, retrospective, multicenter data from routine clinical care, extracted from the electronic medical records of outpatients followed in the nephrology unit of the Hospices Civils de Lyon (HCL) between 01/01/2016 and 31/12/2018. The Photo-Graphe 3 study is a prospective, national, observational, multicenter cohort conducted in France between October 2010 and April 2013. The study involved 62 nephrologists across 20 regions of France. The study aimed to characterize mineral and bone metabolism in patients with stage 4 or 5 CKD who were not receiving dialysis. The HCL dataset, initially used for the first external validation, was incorporated into the training set to enhance the model’s performance [16]. Data from participants who met al.l of the following inclusion criteria: (i) older than 18 years, (ii) clinically diagnosed with stage 4–5 CKD (eGFR ≤ 30 ml/min/1.73 m²), (iii) not on dialysis or having undergone renal transplantation, and (iv) having a known vital status after 2 years of follow-up were included in the training dataset [15, 16]. The combined training dataset was used to retrain the Bayesian network, while retaining the seven variables previously identified as the most informative predictors of 2-year all-cause mortality [15]. External validation dataset The external validation dataset was composed of retrospective, routine care and multicentric data from outpatients of CERRENE, France (Cercle de Recherche, de Réflexion et d’Etudes en Néphrologie), between 01/01/2020 and 31/01/2022, that met all of the inclusion criteria as mentioned above for the training dataset. From the date of study entry, patients were scheduled for 4 follow-up consultations, with 6-month intervals between each follow-up. Patient data were recorded on paper by blinded data collectors - experienced nephrologists of CERRENE. The paper form used to collect data is described in Supplemental Fig. 2. To ensure data accuracy and avoid entry bias, data were manually entered into digital format using a double-entry process. Each data was entered into separate Excel files, which were then compared to identify discrepancies. Any discrepancies were resolved by retrieving the correct values from the original paper forms. The database was subsequently cleaned and categorized in the same manner as the training dataset for the 2-year all-cause mortality prediction tool. Due to the anonymous nature of data collection, the vital status at 2 years could not be verified retrospectively through national records for patients who were no longer followed at the nephrology center, such as those who had transferred to another facility, and were therefore excluded from the validation dataset. Data from patients who did not attend intermediate consultations at 6, 12, or 18 months but whose vital status was known at 24 months were included in the final analysis. Statistical analysis Internal and external validation performance assessment The study was conducted in accordance with the updated version of the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis and Diagnosis (TRIPOD) guidelines, for prediction models using regression or machine learning methods [18]. The checklist is provided in Supplemental Fig. 3. The internal performance of the 2-year all-cause mortality prediction tool was assessed through cross-validation using the following metrics: AUC-ROC, accuracy (%), sensitivity (%), and specificity (%). Similarly, in the external validation, the discriminatory ability of the prediction tool was evaluated using the same metrics. RapidMiner Studio® version 10.3 was used to plot the receiver operating characteristic (ROC) curve and perform external validation [19]. Details of this process are provided in Supplemental Fig. 4. The agreement between the predicted and observed outcomes in the external validation was assessed using a calibration plot. The calibration plot was generated using the PresenceAbsence package version 1.1.11 in R version 4.5.1 with default settings, except for N.bins = 8 [20]. The calibration curve was generated using default setting loess regression in R version 4.5.1 [21]. The code for this procedure is provided in Supplemental Table 1. To address the issue of an imbalanced dataset, we identified an optimal cut-off point using the highest Youden index. The highest Youden index provides the most balanced trade-off between sensitivity and specificity [22]. The interpretation of mortality probability calculations at 2 years was as follows: if the mortality probability of a patient was equal to or greater than the cut-off point, the patient was predicted to be in the deceased group within 2 years. Missing data Missing data in both the training and external validation datasets were imputed using a random forest model. The missing data imputation process was performed in RapidMiner Studio® version 10.3 using the “Impute Missing Value” operator (Supplemental Fig. 4). Minimum sample size calculations The minimum sample size for the validation dataset was calculated using the method proposed by Riley et al. to determine the minimum sample size required for an external validation study of a binary prediction tool [23]. The calculation was performed using the pmvalsampsize package version 0.1.0 in R version 4.5.1 with default settings, except for oeciwidth = 0.2 and lpnorm = c(-0.0535, 1.5498) [24]. For the arguments required in the calculation function, we used the results from the internal validation of the Bayesian network-based 2-year all-cause mortality prediction tool in the development study. Specifically, we anticipated a c-statistic of 0.81 and used the linear predicted values calculated from the predicted probabilities in the development study (-0.0535, 1.5498) [15]. The event prevalence in the population was estimated based on the prevalence observed in our previous external validation, which was 0.17 [16]. The R code for this calculation is provided in Supplemental Table 2. Statistical test Quantitative variables are expressed as mean ± standard deviation (SD) if normally distributed; otherwise, as median and interquartile range (IQR). Discretized quantitative variables and qualitative variables were expressed as percentages. The Student’s t-test was used to compare groups with normal distribution; otherwise, the Mann-Whitney test was used. Distribution of discretized quantitative variables and of qualitative variables were compared using the Chi-square test. A p value < 0.05 was considered significant. These calculations were performed using MedCalc® software version 23.0.2 [25]. Results Training dataset The training dataset included data from 1,061 eligible patients, comprising 534 patients from Photo-Graphe 3 and 527 patients from HCL. Of these, 225 patients (21.2%) were deceased within 2 years of follow-up. The flow chart illustrating the construction of the training dataset is presented in Fig. 1. Flow chart of the training dataset and the external validation dataset Among the seven variables included in the prediction model, there are significant differences between the deceased group and the survivors in terms of CV history, smoking status, PTH levels, and the use of ESA. In particular, patients in the deceased group were significantly older than those in the survivor group (median 80.0 years vs. 70.3 years, respectively, p < 0.001). The ratio of current smokers was significantly higher in the deceased group (11.6% vs. 11.3%, p = 0.04). The percentage of patients with a CV history in the deceased group was nearly double that of the survivor group (60.9% vs. 35.9%, p < 0.001). PTH levels were significantly higher in the deceased group (median 133 vs. 109 pg/mL, p = 0.008). More patients were treated with ESA in the deceased group than in the survivor group (37.1% vs. 22.0%, p < 0.001). The characteristics of the patients included in the training dataset are summarized in Table 1. Table 1Characteristics of the training datasetVariableTotal(N = 1061)Deceased group(N = 225)Survivor group(N = 836)Sub-group comparisonMedianIQRDistribution %MedianDistribution %MedianDistribution %p-value Age (year) 72.3 62.0–81.0 80.0 70.3 < 0.001*Men57.1%65.854.80.003**BMI (kg/m2)27.123.9–31.126.327.20.02*SBP (mmHg)140128–1521401400.33*DBP (mmHg)75.067.0–83.073.076.00.002* Current smoker 11.4% 11.6 11.3 0.04** CV history 41.1% 60.9 35.9 < 0.001** eGFR (mL/min/1.73m2)21.917.4–26.020.322.4< 0.001*Hemoglobin (g/dL)11.910.8–12.911.711.90.05* Ferritin (µg/L) 139 83.0–228 151 136 0.08* LDL-C (g/L)0.930.67–1.220.890.930.48*Calcemia (mmol/L)2.312.23–2.402.302.320.01*Phosphorus (mmol/L)1.251.10–1.411.301.24< 0.001* PTH (pg/mL) 112 67.7–187 133 109 0.008* 25-OH vitamin D (nmol/L) 70.0 44.0–93.0 65.0 70.5 0.19* Bicarbonate (mmol/L)23.021.0–26.024.023.00.10* Use of ESA 25.2% 37.1 22.0 < 0.001** Use of VKA30.2%30.029.40.29*** Mann-Whitney test** Chi-square testAbbreviation: CV: Cardiovascular; DBP: Diastolic blood pressure; eGFR: Estimated glomerular filtration rate; ESA: Erythropoiesis-stimulating agent; LDL-C: Low-density lipoprotein-cholesterol; PTH: Parathyroid hormone; SBP: Systolic blood pressure; VKA: Vitamin K antagonist oral anticoagulantThe variables written in bold and italic are the predictors included in the 2-year all-cause mortality prediction tool Characteristics of the training dataset * Mann-Whitney test ** Chi-square test Abbreviation: CV: Cardiovascular; DBP: Diastolic blood pressure; eGFR: Estimated glomerular filtration rate; ESA: Erythropoiesis-stimulating agent; LDL-C: Low-density lipoprotein-cholesterol; PTH: Parathyroid hormone; SBP: Systolic blood pressure; VKA: Vitamin K antagonist oral anticoagulant The variables written in bold and italic are the predictors included in the 2-year all-cause mortality prediction tool External validation dataset The minimum calculated sample size for external validation was n = 1,878 patients, with a minimum number of events corresponding to an event prevalence of 0.17, corresponding to 320 events. After the exclusion of 44 patients lost to follow-up (9.71%), 409 patients were included in the external validation dataset (Fig. 1). The percentage of missing data in the CERRENE dataset was 14.7%. The proportion of deceased patients in the external validation dataset was lower, though not significantly, than in the training dataset (17.6% vs. 21.2%, respectively). Patients in the external validation dataset had a slightly, yet statistically significant, higher eGFR compared to those in the training dataset (23.0 vs. 21.9 ml/min/1.73 m²; p = 0.03). Among the seven predictors included in the prediction tool, there were significant differences between the training dataset and the external validation dataset in terms of age, use of ESA, and 25-OH vitamin D levels. The comparison between the training and external validation datasets is presented in Supplemental Table 3. In the external validation dataset, the percentages of deceased patients after 6, 12, 18, and 24 months of follow-up were 4.89%, 9.54%, 12.71%, and 17.6%, respectively. The majority of patients had comorbidities, with 95.6% having hypertension and 59.7% having type 2 diabetes. Vascular and diabetic nephropathy were the most common etiologies of CKD in the external validation dataset (40.8% and 31.5%, respectively). As in the training dataset, deceased patients in the external validation dataset were significantly older and had higher proportions of patients with a CV history and those treated with ESA compared to the survivor group. However, the proportion of current smokers in the deceased group was significantly lower than in the survivor group (4.29% vs. 9.43%, p = 0.03), which contrasts with the training dataset. There was no significant difference in PTH levels between the two subgroups in the CERRENE dataset. The characteristics of the external validation dataset are described in Table 2. Table 2Characteristics of the external validation datasetVariablesTotal(N = 409)Deceased group(N = 72)Survivor group(N = 337)Sub-group comparisonMedianIQR*Distributrion %MedianDistribution %MedianDistribution %p-valueAge (year)77.569.8 to 82.981.176.5< 0.001Men64.866.764.80.09**BMI (kg/m2)28.725.2 to 32.226.628.70.01*SBP (mmHg)143130 to 1551441430.73*DBP (mmHg)76.067.0 to 84.073770.12* Current smoker 8.10 4.29 9.43 0.03** CV history 46.0 68.1 41.5 < 0.001** eGFR (mL/min/1.73m2)23.018.0 to 26.32024< 0.001*Hemoglobin (g/dL)12.011.0 to 13.211.512.20.02* Ferritin (µg/L) 156 77.0 to 259 157 156 0.46* LDL-C (g/L)0.940.65 to 1.261.070.930.26*Calcemia (mmol/L)2.332.25 to 2.422.322.330.94*Phosphorus (mmol/L)1.161.06 to 1.321.171.160.33* PTH (pg/mL) 115 71.0 to 168 125 113 0.30* 25-OH vitamin D (nmol/L) 58.0 37.0 to 89.3 50 58.3 0.29* Bicarbonate (mmol/L)24.022.0 to 26.025240.02* ESA 9.30 23.6 6.25 < 0.001** VKA31.847.228.50.002*** Mann-Whitney test** Chi-square testAbbreviation: CV: Cardiovascular; DBP: Diastolic blood pressure; eGFR: Estimated glomerular filtration rate; ESA: Erythropoiesis-stimulating agent; LDL-C: Low-density lipoprotein-cholesterol; PTH: Parathyroid hormone; SBP: Systolic blood pressure; VKA: Vitamin K antagonist oral anticoagulantThe variables written in bold and italic are the predictors included in the 2-year all-cause mortality prediction tool Characteristics of the external validation dataset * Mann-Whitney test ** Chi-square test Abbreviation: CV: Cardiovascular; DBP: Diastolic blood pressure; eGFR: Estimated glomerular filtration rate; ESA: Erythropoiesis-stimulating agent; LDL-C: Low-density lipoprotein-cholesterol; PTH: Parathyroid hormone; SBP: Systolic blood pressure; VKA: Vitamin K antagonist oral anticoagulant The variables written in bold and italic are the predictors included in the 2-year all-cause mortality prediction tool Predictive performance assessment The AUC-ROC and specificity of the external validation were slightly lower than those in internal validation, whereas accuracy and sensitivity were slightly higher. In the internal 10-fold cross-validation, the prediction tool showed an accuracy of 77.2% (SD 2.75), an AUC-ROC of 0.76 (SD 0.05), sensitivity of 47.1% (SD 12.1), and specificity of 85.3% (SD 2.26). The optimal cut-off point selected was 0.347, corresponding to the highest Youden index (0.37). The Youden index and cut-off point are illustrated in the ROC plot in Fig. 2. In the external validation, the prediction tool showed an AUC-ROC of 0.74. At the optimal cut-off point, it yielded an accuracy of 77.8%, sensitivity of 54.2%, and specificity of 82.3%. The external validation performance of the prediction tool is detailed in Table 3. The calibration plot of the external validation is shown in Fig. 3, with Brier score = 0.132. Receiver operating characteristic curve showing the performance of the prediction tool at classification cut-off. The solid green line represents sensitivity. The solid yellow line represents the Youden index. The solid orange line represents the cut-off point. The highest Youden index was 0.37 and its corresponding cut-off point was 0.347. The corresponding sensitivity was 0.54 and 1-specificity was 0.17 (i.e. specificity 0.83) Table 3Internal and external predictive performanceInternal 10-fold cross validationExternal validationAccuracy %77.2 ± 2.7577.8AUC-ROC0.76 ± 0.050.74Sensitivity %47.1.±12.154.2Specificity %85.3 ± 2.2682.3Abbreviation: AUC-ROC: area under the receiver operating characteristic curve Internal and external predictive performance Calibration plot of the external validation Discussion The accuracy of the external validation performance of the Bayesian Network-based 2-year all-cause mortality prediction tool was enhanced by augmenting the training dataset and applying an optimal cut-off point. Specifically, the accuracy reported in this study is 77.8%, reflecting a 22.3% increase compared to previous external validation results (63.6%) [16]. The Bayesian network-based 2-year all-cause mortality prediction tool demonstrated acceptable calibration overall. Most prediction tools developed for patients with CKD focus on predicting disease progression, with outcomes such as renal replacement therapy initiation, decline in eGFR, or an increase in serum creatinine. A few all-cause mortality prediction tools have been developed in recent years. An analysis showed that only 14% of studies on prediction tools for patients with CKD referenced the TRIPOD guidelines after their release in 2015 [26]. Additionally, only 28% of all-cause mortality prediction tools for patients with CKD were both developed and externally validated within the same study [26]. The prediction tools developed by Alex et al. [9], Ning Li et al. [10], Ping Liu et al. [11], and Kanda et al. [12] were designed to predict mortality but targeted patients with CKD at different stages than our study. The prediction tool by Bansal et al. aimed to predict 5-year mortality risk in patients with CKD stages 4–5 but was restricted to individuals aged ≥ 65 years [13]. This tool demonstrated weak discriminatory ability [13]. The prediction tool by Grams et al. focused on the same CKD population as our study (CKD stages 4–5) [14]. It simultaneously assessed the risks of end stage kidney disease, nonfatal cardiovascular events, and death [14]. However, mortality prediction results were poorly reported in the original article [14]. Additionally, the study by Grams et al. lacked external validation and did not refer to the TRIPOD guidelines [14]. This tool included nine variables: age, sex, race, cardiovascular history, smoking status, systolic blood pressure, diabetes mellitus, eGFR, and urine albumin-creatinine ratio. However, the study did not provide detailed descriptions of variable selection [14]. This study reproduced the previously reported Bayesian network-based 2-year all-cause mortality prediction tool using a larger training dataset, followed the TRIPOD guidelines, and was then externally validated with an independent dataset. Both internal and external predictive performances were satisfactory. However, compared to the previously reported internal validation performance, the internal validation in this study showed lower metrics [15]. The training dataset used herein to build the Bayesian network is larger than that of the previous development study (n = 1,061 vs. n = 534), but the event ratio is lower (21.2% vs. 25.1%). This higher event ratio and the use of synthetic data in or previously reported prediction tool may contribute to better overall internal performance. In comparison to the previously reported external validation, the prediction tool demonstrated better performance across nearly all metrics, except for sensitivity [16]. There were discrepancies in sensitivity and specificity between the internal and external validation. These differences may be attributed to variations in the predicted probability distributions and the proportion of deaths between the training and external validation datasets. To improve the performance of the prediction tool on an imbalanced dataset, previous studies used the Synthetic Minority Over-sampling Technique (SMOTE) proposed by Chawla et al. [27] to generate synthetic training data. However, reports have highlighted potential harms of using class imbalance correction techniques on predictive performance [28]. Instead, selecting an optimal cut-off point to maximize predictive performance metrics on an imbalanced dataset has been found to achieve good predictive accuracy [28]. Therefore, we did not generate a synthetic dataset to balance the minority class in our training population. The information obtained from the prediction tool can assist physicians in delivering more accurate and detailed information to patients, empowering them to engage in the decision-making process. Thus, an accurate prediction tool can help physicians provide better personalized management for patients with CKD, thereby improving clinical outcomes. This 2-year all-cause mortality prediction tool includes seven predictors: age, cardiovascular history, smoking status, levels of ferritin, PTH, 25-OH vitamin D, and use of ESA, all of which are easy to collect in clinical practice. This may facilitate the integration of this prediction tool into clinical practice. The seven variables were selected from an initial set of 26 candidate predictors, which were identified based on the literature and expert opinion. Unfortunately, proteinuria was not recorded in the initial development dataset (dataset of Photo-Graphe 3 study). However, proteinuria is likely to be more strongly associated with clinical outcomes in earlier stages of CKD. Missing data remains a significant challenge in clinical practice. The Bayesian network is a mathematical model that has shown good effectiveness in handling missing or censored data in health records [29]. Furthermore, an analysis by Arora et al. highlighted the advantages of Bayesian networks over regression-based approaches, including (i) the creation of network that clearly convey relationships between variables, (ii) the ability to utilize Bayes’s theorem for estimating individual-level risk, and (iii) their straightforward adaptability into decision models [30]. Although external validation was performed using an independent dataset, it is important to note that these data were also collected in France. It should be noted that the low percentage of missing data in the training and validation databases supports the high quality of the data. Additionally, the total sample size and number of events in the external validation dataset fell short of the minimum sample size calculated for external validation (409 vs. 1,878 patients and 72 vs. 320 events). The insufficient sample size and number of events in our external validation population may explain the wide confidence intervals observed in the calibration analysis and less satisfactory sensitivity. The calibration plot also showed that the prediction tool tends to overestimate risk, particularly for patients in the intermediate risk range. The training dataset had a median age of 72.3 years, with 22% patients younger than 60 years of age. This imbalance may limit the generalizability of the prediction tool for younger patients. Before it can be used in clinical practice on a global scale, it must undergo national and international external validations in larger and more diverse populations, which are currently underway. Moreover, the clinical impact of this Bayesian network-based 2-year all-cause mortality prediction tool should be evaluated before it can be widely implemented in clinical practice. The prediction tool is available for research purposes at the following address: CKD-M2 prediction tool. Conclusion We developed and externally validated a Bayesian network–based model to predict 2-year all-cause mortality in patients with stage 4–5 CKD using seven readily accessible clinical predictors. The model demonstrated satisfactory performance, with improved accuracy after expanding the training dataset. However, the prediction tool tended to overestimate risk among patients in the intermediate risk range. Although these findings support its potential clinical utility, further external validation in larger and more diverse populations is underway and will be essential before broader implementation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We would like to thank Professor Thierry Hannedouche, M.D., for his careful review of this article.We would like to thank nephrologists from CERRENE; Claire Borni Duval M.D., and Clotilde Muller, M.D., who carefully recorded the data of CKD patients. Abbreviations Area under the receiver operating characteristic curve Cercle de Recherche, de Réflexion et d’Etudes en Néphrologie Chronic kidney disease Cardiovascular Estimated glomerular filtration rate Erythropoiesis-stimulating agents Hospices Civils de Lyon Interquartile range Parathyroid hormone Receiver operating characteristic Standard deviation Synthetic Minority Over-sampling Technique Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis and Diagnosis guideline Author contributions Dung Nu Thuy Tran and Jean-Pierre Fauvel: Methodology, data transformation, data analysis, manuscript preparation.Michel Ducher and Jean-Pierre Fauvel: Conceptualization, methodology.Denis Fouque, Yves Dimitrov, François Chantrel, Peggy Perrin and Zead Tubail: data collection.All the authors have read and approved the manuscript. Funding The first author, Dr. Tran has received a scholarship from the Société Francophone de Dialyse et de Transplantation (SFNDT) for her doctoral thesis. The sponsors had no influence on the data collection, statistical analysis and writing of the paper. Partial financial support was received from research grants of GlaxoSmithKline, AstraZeneca, Astellas and Servier. The sponsors had no influence on the data collection, statistical analysis and writing of the paper. This study was partly funded by the MESSIDORE research fund of the French National Institute of Health & Medical Research (INSERM). The sponsors had no influence on the data collection, statistical analysis and writing of the paper. Data availability The datasets generated during or analyzed for the present study are available from the corresponding author upon reasonable request. Declarations Ethical considerations and informed consent The study was carried out in accordance with the French law on retrospective studies of data recorded in the hospital system. The information letter sent to the patient’s address was reviewed and endorsed by the local ethics committee. The patient or the patient’s family (if the patient is deceased) had one month to express their opposition to the use of their personal data. This study was conducted in strict compliance with the Declaration of Helsinki and was approved on July 28, 2022 by the Scientific and Ethical Committee of the Hospices Civils de Lyon (number 22 835) and declared to the national data protection commission (number 22 5835). Consent for publication Patients or their legal representatives have given their informed consent for publication. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets generated during or analyzed for the present study are available from the corresponding author upon reasonable request.