Comparative prognostic accuracy of MDRD and CKD-EPI equations in Jordanian diabetics: A machine learning-enhanced longitudinal study

Abstract

Background:

Type II diabetes mellitus (DM) is a leading cause of chronic kidney disease (CKD), increased risks of cardiovascular events, and all-cause mortality. Even mild reductions in glomerular filtration rate (GFR) in people with DM confer higher risks of death and adverse cardiovascular outcomes, underscoring the importance of accurate kidney function assessment for risk stratification and management.

Objective:

This study aimed to compare the prognostic accuracy of the Modification of Diet in Renal Disease (MDRD) and CKD-Epidemiology Collaboration (CKD-EPI) equations for predicting all-cause mortality in a longitudinal cohort of Jordanian patients with DM type II.

Design:

This is a cohort retrospective longitudinal study.

Methods:

An analysis of electronic health records was performed on data from 1884 patients with DM, utilizing both traditional statistical approaches and machine learning techniques.

Results:

The median follow-up was 38 months. 63.2% and 56.8% of the 1884 diabetic patients were diagnosed with CKD using the MDRD and CKD-EPI formulas, respectively. CKD at admission was linked to a significantly higher risk of death, with adjusted odds ratios of 3.29 for MDRD and 3.39 for CKD-EPI. The CKD-EPI formula emerged as the most critical factor in predicting mortality, particularly when analyzed using the Random Forest (RF) model.

Conclusion:

The RF model effectively demonstrated the superiority of capturing relevant patterns in the feature set when using the CKD-EPI equation compared with the MDRD equation in diabetic patients. Clinicians should use the CKD-EPI equation to predict CKD and death in low-resource settings only after thorough local validation in diabetic populations.

Plain language summary

MDRD or CKD-EPI: Which kidney formula best protects diabetic patients?

Why was this study done? Clinicians often use different mathematical formulas (called “equations”) to check how well a patient’s kidneys are working. Two common formulas are “MDRD” and “CKD-EPI.” This study was done in Jordan to find out which of these two formulas is better at predicting the risk of death for patients with type 2 diabetes.

What did the researchers do? The researchers looked back at the electronic health records of 1,884 Jordanian patients with type 2 diabetes. They used both the MDRD and CKD-EPI formulas to see which patients had chronic kidney disease (CKD). They then used statistics and advanced computer analysis (machine learning) to see which formula was more accurate in predicting which patients passed away over a five-year period.

What did the study find? The study revealed kidney disease in over half of the patients. The number changed slightly depending on which formula was used. Having kidney disease upon initial hospital admission significantly increased the risk of death. The computer analysis showed that the CKD-EPI formula was the best single tool for predicting mortality. Overall, the CKD-EPI formula was more reliable than the MDRD formula for finding patterns in the data that predict death in patients with diabetic patients.

What does this mean for patients? This research suggests that using the CKD-EPI formula gives clinicians a more accurate picture of a diabetic patient’s kidney health and their future risk. This study can help clinicians and patients make better-informed decisions about care and treatment plans.

Keywords

chronic kidney disease MDRD CKD-EPI random forest model electronic health records and mortality

Introduction

The global prevalence of diabetes mellitus (DM) type II stands at an estimated 422 million adults and is the 8th leading cause of death.¹ It is noteworthy that the prevalence of type 2 diabetes has surged more rapidly in low and middle-income countries (LMICs) when compared to high-income countries (HICs) on a global scale.^1,2 Diabetes leads to a range of complications, including cardiovascular disease, retinopathy, neuropathy, and nephropathy. Chronic kidney disease (CKD) is progressive, advances to end-stage renal disease (ESRD), and increases the risk of death among patients with DM. The prevention and screening of CKD begin with identifying its epidemiology and the distribution of risk factors.³ For Jordanian patients with diabetes, stratifying their risk of CKD leads to more specific, effective treatment and better results. For patients with DM type II, intrinsic factors such as obesity, hypertension (HTN), and vascular aging could affect the accuracy of diagnosing CKD.^4,5

Renal function is typically assessed through the estimated glomerular filtration rate (eGFR), which is calculated using creatinine-based equations.^6,7 Validated noninvasive risk equations for eGFR present an appealing option for CKD screening in diabetic patients, facilitating the identification of individuals at risk.^8,9 The Modification of Diet in Renal Disease (MDRD) equation is commonly used to predict CKD in patients with identified risk factors. Using the MDRD equation, the CKD prevalence was estimated to be around 8% among outpatients with chronic illnesses.⁴ Alternatively, the CKD-Epidemiology Collaboration (CKD-EPI) equation, which incorporates age, gender, and ethnicity, shows greater accuracy. Nevertheless, the CKD-EPI equation has never been evaluated against the MDRD equation in Jordanian patients with diabetes, as in LMICs, for predicting health-related outcomes. The diabetic population in Jordan has distinct demographic and genetic traits that may impact kidney function and the prevalence of related diseases. Given the high rates of diabetes and diabetic kidney disease in this population, we validated two GFR-estimating equations. Factors such as ethnicity, lifestyle, and environmental influences can affect the performance of equations such as MDRD and CKD-EPI in this specific context. Furthermore, no studies have yet assessed the validity of these equations within the Jordanian diabetic population, highlighting the need to determine which equation is most suitable for clinical practice in this group.^3,10–12

It is worth mentioning that published Jordanian studies have reported the clinical profiles and prognoses of populations at risk, but few specifically address the renal-related outcomes among patients with chronic illnesses.^2,4,13,14 It is important to accurately identify renal parameters, such as eGFR and serum creatinine levels, from extensive datasets to guide assessment and prognosis.⁸ Although the CKD-EPI and MDRD equations were developed and validated in HICs,¹⁵ their accuracy within the Jordanian context remains uncertain. Therefore, dedicated studies are essential to elucidating the potential consequences of applying non-validated equations, which may include clinical misdiagnosis and flawed public health planning.^16–18 By analyzing hidden patterns within Electronic Health Records (EHRs), machine learning (ML) algorithms could enhance the predictive accuracy of standard equations (MDRD and CKD-EPI), particularly in LMICs^3,10 where their performance compared to developed nations requires further validation.3 While the CKD-EPI equation is recognized for its improved accuracy, the MDRD formula is still routinely used in some centers in Jordan. However, to the best of our knowledge, no studies have specifically evaluated the performance of these equations in our population. As a result, we do not yet know which equation is more suitable for accurately estimating the GFR in Jordanian patients. This study aims to fill that gap and provide valuable insights for clinical practice. The objective of this study was to examine the performance of using ML models and statistical analysis methods in EHRs for Jordanian patients with DM.

This study addresses a critical gap in CKD management by comparing the performance of CKD-EPI and MDRD equations in predicting mortality among diabetic patients in a middle-income setting. While both equations are widely used for CKD diagnosis, evidence on their prognostic utility in low-resource populations is limited.³ CKD staging and prognosis have traditionally relied on eGFR equations such as the CKD-EPI formula, which has been shown to provide more accurate risk estimation compared with other equations in diverse populations.¹⁰ Previous studies have also demonstrated the potential of ML methods, such as random forest (RF) models, to enhance the prediction of CKD progression and related outcomes, sometimes outperforming nephrologists in forecasting end-stage kidney disease (ESKD).¹⁰

However, it remains unclear whether the integration of ML approaches can further improve the predictive accuracy of eGFR equations for mortality risk and how such models may aid in refining CKD staging and prognosis. Therefore, this study aimed to evaluate and compare the performance of the CKD-EPI equation using both ML-based and traditional statistical methods in predicting mortality among CKD patients. This study underscores the importance of validating globally recognized equations in diverse populations to ensure equitable healthcare delivery.

Study design

This was a national, observational, retrospective study performed between 2011 and 2021 at the largest public hospital in the capital serving approximately 3 million residents. A total of 1884 patients with DM-II in outpatient departments were enrolled with follow-up up to 5 years after inclusion. Renal data were obtained from the EHRs that integrate data across public, private, and military hospitals. This study followed the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines (Supplemental material).

Sample and population

The study used nationwide de-identified EHRs containing longitudinal health data of patients visiting outpatient departments for DM check-ups from January 2011 to December 2021 in public Jordanian hospitals. It enrolled 12,000 patients with chronic medical diseases, identified through searching, aged over 18, and had a clinic visit in the public hospital that serves many geographical areas in Jordan. Duplication of data was checked, and the number of patients was reduced from 12,000 to 7000 patients diagnosed with chronic medical diseases. Patients were screened according to the eligibility criteria, resulting in the inclusion of 1884 adult patients with type 2 DM (DM-II), each having at least two consecutive serum creatinine measurements. A total of 1884 eligible patients were included in the final analysis. The exclusion criteria were consistent and included patients with defined CKD and/or receiving dialysis treatment.

Instruments

The author communicated with Hakeem to obtain ethical institutional review board approval from the public (the Ministry of Health) to extract data from the patients’ EHRs that include demographic and clinical variables.

Measurements of the glomerular filtration rate

Demographic data, including gender, age, and duration of diabetes, and clinical data, including history of hypertension and serum creatinine (two stable readings of serum creatinine within 90 days), were extracted from the EHS. The CKD equations are formulas used to estimate the GFR, which is a measure of how well the kidneys are functioning, including the MDRD and EPI-CKD (Chronic Kidney Disease Epidemiology Collaboration) equations.

eGFR Equations

eGFR(MDRD) = 175 × (Scr)^(−1.154) × (Age)^(−0.203) × (0.742 if female).

eGFR(CKD-EPI) = 141 × min(Scr/k, 1)^α × max(Scr/k, 1)^(−1.209) × 0.993^(Age) × (1.018 if female).

Where: k = 0.7 for females and 0.9 for males; α = −0.329 for females and −0.411 for males. The stages of CKD were determined by eGFR and defined according to the National Kidney Foundation Disease Outcome Initiative (CKF-KDOQI) guidelines. Stage 1 is when eGFR = ⩾90, stage 2 is when eGFR = 60–<90, stage 3 is when eGFR = 30–<60, stage 4 is when eGFR = 15–<30 and stage 5 is when eGFR = <15 mL/min/1.73 m².¹⁹

Health outcomes

The primary end point was the occurrence of CKD at any stage using the MDRD and CKD-EPI equations. According to worldwide clinical recommendations, CKD is defined as an eGFR less than 60 mL/min/1.73 m² (stage 3–5). According to each equation, eGFR was calculated from electronic medical record serum creatinine measurements. Participants were classified as having CKD or not based on MDRD and CKD-EPI estimates individually for each equation.

The secondary end point was all-cause mortality during the follow-up period. From hospital records and medical record dates, the mortality status was determined. Regardless of cause, mortality was regarded as a binary outcome (died or not) for analysis.

Statistical analysis

Managing and preprocessing data

Initially, a huge dataset of about 7000 hospitalized patients with chronic illness was collected. We cleaned the data by getting rid of duplicates, missing entries, and wrong entries, and we standardized or changed the variables as needed. After applying the inclusion criteria, the final analytical sample comprised 1884 patients with DM.

Descriptive statistics

To summarize continuous variables, we used means and standard deviations. To summarize categorical variables, we used frequencies and percentages. We used the MDRD and CKD-EPI equations to figure out the estimated GFR values individually.

Group comparisons between MDRD and CKD-EPI

Since both equations were applied to the same cohort of patients, comparisons were made based on the eGFR values produced by each equation, not on patient characteristics, which are inherently identical across equations.

Conventional statistical techniques

Data were analyzed using IBM SPSS Statistics, version 21.0 (IBM Corp., Armonk, NY, USA). A p value of less than 0.05 was considered significant. To analyze variations in eGFR estimates based on patient variables, independent t-tests, Chi-square tests, and Pearson correlation coefficients were utilized in accordance with the measurement level. Pearson correlations examined the association between continuous variables like age and diabetes duration and categorical predictors such as gender, HTN, congestive heart failure, retinopathy, and eGFR for each equation.

Modified Bland–Altman analysis was used, and the Kappa statistics were used to see if the equations agreed with each other. Crosstabulation was employed to analyze concordance among CKD phases. Kaplan–Meier survival curves and the log-rank test were employed to assess the risk of death between different types of CKD.

Analyzing machine learning

Preparing data for ML

We used Python software (version 3.10; Python Software Foundation, Beaverton, OR, USA) to model ML. Before training, missing values were filled in using median imputation, and continuous predictors were normalized. Using stratified sampling to keep the class distribution the same, the data were split into training (80%) and testing (20%) subsets. Five-fold cross-validation within the training set was used to test how well the model might be used in other situations.

Building the RF model

We utilized an RF classifier (sklearn.ensemble) to guess who would die. Mortality (died vs survived) was defined as the binary outcome variable. CKD status was determined separately using the CKD-EPI and MDRD equations, applying an eGFR <60 mL/min/1.73 m² to create two independent binary predictors. Both CKD-EPI-defined CKD and MDRD-defined CKD variables were entered simultaneously into the ML model together with demographic and clinical covariates known to influence mortality, including age, sex, DM duration, neuropathy, hypertension, myocardial infarction, nephropathy, retinopathy, and congestive heart failure. To compare the importance of the CKD estimation formula in the prediction of mortality, we trained an RF model (observed prediction error = 0.063) and used permutation scores to evaluate the importance of each predictor (Figure 3). Feature importance metrics were used to compare the relative contribution of CKD classification by each equation to mortality prediction within the same multivariable model. A grid search was used to find the best values for the hyperparameters. This included the number of trees (n_estimators), the depth of the trees (max_depth), the lowest number of samples needed to divide a node (min_samples_split), and the size of the leaves (min_samples_leaf). The best setup was:

n_estimators = 1000

max_depth = 7

min_samples_split = 20

min_samples_leaf = 5

random_state = 24

This setup struck a good compromise between performance and the risk of overfitting.

Checking the model

We looked at the model’s accuracy, AUC-ROC, precision, recall, and F1-score to see how well it worked. The independent test set, which was not used during training or tuning, was employed to get an impartial estimate of predicted performance for the final evaluation.

Results

Patients with DM were admitted for diverse related reasons, including heart failure, coronary artery disease, and stroke. The characteristics of 1884 patients with DM are presented in Table 1. The mean age was 61.8 years, and 47% were female. Jordanian nationality was represented by 1823 patients (96.8%). The median follow-up time was 38 months for the 1884 patients with DM, and the percentages of patients with CKD diagnoses were 63.2% and 56.8% using the MDRD and CKD-EPI formulas, respectively. The univariate correlation analysis revealed that the MDRD and CKD-EPI mean scores decreased significantly with aging (r = −0.45 and −0.53, p value < 0.01, respectively) and duration of DM (r = −0.105 and −0.124, p value < 0.01, respectively). Therefore, the correlation coefficients (r = −0.105 for MDRD and r = −0.124 for CKD-EPI; both p < 0.01) indicate that as diabetes duration increases, estimated kidney function decreases, and this relationship is statistically significant for both equations. The estimated GFR via MDRD and CKD-EPI was significantly lower among males compared to females, patients with hypertension, congestive heart failure, and retinopathy (Table 1).

Table 1.

Differences in CKD estimation equations by sample characteristics (n = 1884).

Variables	n (%) or mean ± SD	MDRD equation, mean ± SD (mL/min/1.73 m²) or r²	CKD-EPI equation, mean ± SD (mL/min/1.73 m²) or r²
Age (mean (SD))	61.76 (14.98)	−0.148**	−0.506
DM duration (years)	3.0 ± 3.1	R = −0.105**	R = −0.124**
Gender
Male	999 (53.0%)	56.3 ± 46.6**	52.4 ± 35.7**
Female	885 (47%)	71.2 ± 63.6	61.4 ± 41.4
HTN
Yes	1029 (54.6%)	52.1 ± 39**	48.6 ± 32.5**
No	855 (45.4%)	76.8 ± 67.8	66.3 ± 43.2
Congestive heart failure
Yes (mL/min/1.73 m²)	99 (5.3%)	42.8 ± 26.3**	40.5 ± 23.3**
No	1785 (94.7%)	65.2 ± 47.5	59.0 ± 34.7
Neuropathy
Yes (mL/min/1.73 m²)	82 (4.4%)	84.6 ± 61.5**	71 ± 37.8**
No	1802 (95.6%)	63.1 ± 45.9	57.5 ± 34.3
Retinopathy
Yes (mL/min/1.73 m²)	43 (2.3%)	40.8 ± 26.6**	40 ± 26.4**
No	1841 (97.7%)	64.6 ± 47.1	58.4 ± 34.6

SD: standard deviation; MDRD: Modification of Diet in Renal Disease equation; CKD-EPI: Chronic Kidney Disease Epidemiology Collaboration equation; DM: diabetes mellitus; HTN: hypertension; CHF: congestive heart failure; CKD: chronic kidney disease.

Pearson correlation coefficients (r) were used for continuous variables (age, DM duration). Independent samples t-tests were used to compare mean estimated GFR values (MDRD and CKD-EPI) across categorical variables (gender, HTN, CHF, neuropathy, retinopathy).

p < 0.01.

We evaluated the agreement between the estimated GFR using MDRD and EPI on 29,970 laboratory values. The final cohort was 1884 patients with DM-II with two serum creatinine values within 90 days. The number was reduced to 1795 patients with complete paired data and no one or more time points missing variables between 2011 and 2021. Figure 1 shows the Bland–Altman plot for the agreement between the GFR values estimated using these two formulas. As seen in this figure, there is a strong agreement between the MDRD and EPI formulas for estimating low GFR values. However, the discrepancy between the two formulas increases at larger GFR values (healthier patients) with a systematic trend of underestimation using the MDRD formula. To evaluate the impact of GFR overestimation on the assignment of CKD stage, we evaluated the categorical distribution of stages using both MDRD and EPI formulas (Table 2). As seen in this confusion matrix, there is a strong agreement between both formulas (Kappa = 0.899, p < 0.001). The MDRD formula tends to assign a lower CKD stage in roughly 5.6% of instances. The CKD-EPI formula provides a more accurate estimation of the GFR across a wider range of kidney functions, especially at higher GFR levels.

Figure 1.

Bland–Altman plot for the agreement between MDRD and EPI formulas for estimating GFR.

Table 2.

Crosstab between CKD stages as assigned by MDRD versus EPI formulas.

MDRD stage	EPI stage I	EPI stage II	EPI stage IIIa	EPI stage IIIb	EPI stage IV	EPI stage V	Total
Stage I	6500	355	0	0	0	0	6855
Stage II	521	5270	180	0	0	0	5971
Stage IIIa	0	149	3164	229	0	0	3542
Stage IIIb	0	0	100	3025	400	0	3525
Stage IV	0	0	0	7	3164	526	3697
Stage V	0	0	0	0	0	6380	6380
Total	7021	5774	3444	3261	3564	6906	29,970

CKD: chronic kidney disease; eGFR: estimated glomerular filtration rate; MDRD: Modification of Diet in Renal Disease equation; CKD-EPI: Chronic Kidney Disease Epidemiology Collaboration equation.

CKD stages are defined according to KDIGO eGFR cutoffs. Values represent frequency counts from cross-classification of CKD stages derived using the MDRD and CKD-EPI equations.

The following variables were controlled in the models: age, DM duration, neuropathy, hypertension, congestive heart failure, myocardial infarction, nephropathy, gender, and retinopathy. The diagnosis of CKD during the last admission was associated with a higher risk of death using both the MDRD (adjusted odds ratio (aOR) = 3.29, 95% CI = 1.86–5.8, p < 0.001) and the CKD-EPI formulas (OR = 3.39, 95% CI = 1.9–5.9, p < 0.001). To compare the prognostic value of CKD stage assignment on mortality using MDRD and EPI formulas, we used Kaplan–Meier curves with Log-Rank Chi-square test (Figure 2(a) and (b)). As seen in this figure, both class assignments using either of the formulas provide clear separation in the survival curves of the time-to-death data, with the EPI formula providing slightly better separation. Kaplan–Meier curves (Figure 2(a) and (b)) illustrate that individuals classified as having CKD by either MDRD or CKD-EPI experienced higher mortality compared with those without CKD using the same equation. Because both equations were applied to the same population, the curves appear similar and are not intended to visually demonstrate comparative predictive superiority. The comparison of predictive performance between equations is instead provided through C-statistics and adjusted hazard ratios.

Figure 2.

(a) Kaplan–Meier survival curves of presence or absence of CKD assignment on mortality-MDRD. Log rank = 24,9, p value < 0.001. (b) Kaplan–Meier survival curves of presence or absence of CKD assignment on mortality-CKD-EPI. Log rank = 26,2, p value < 0.001.

RF variable importance was quantified using the Mean Decrease in Accuracy by Permutation, which reflects how much the model’s predictive accuracy declines when a given variable is randomly permuted. Higher values indicate predictors that have a greater impact on model performance. Error bars represent the standard deviation of importance values across multiple permutations, indicating the stability of each predictor’s contribution to the model. Clinically, variables with higher importance scores are more influential predictors of mortality in our cohort. To optimize the RF model for mortality prediction, we performed hyperparameter tuning to improve predictive performance while minimizing overfitting. The final model included 1000 decision trees (n_estimators = 1000), providing greater stability and accuracy. Model complexity was controlled by limiting the maximum tree depth to seven levels (max_depth = 7), ensuring that the algorithm captured meaningful patterns rather than noise. Additional regularization was achieved by requiring at least 20 samples to split an internal node (min_samples_split = 20) and at least 5 samples per terminal node (min_samples_leaf = 5), reducing the likelihood of overly specific or unstable splits. Finally, we set random_state = 24 to ensure that all analyses were reproducible by producing the same sequence of random operations each time the model was run (Figure 3). Figure 3(a) shows the area under the ROC curve of the RF classifier for predicting mortality in the hold-out test set (80%), and Figure 3(b) shows the classification performance matrix. Cutoffs were based on 5-fold cross-validation on the training set, with a model optimized for higher recall and negative predictive value.

Figure 3.

Performance of the RF classifier on the hold-out test set (n = 377). This figure shows (a) the area under the ROC curve of the RF classifier for predicting mortality in the hold-out test set (80%), and (b) the classification performance matrix. Cutoffs were based on 5-fold cross-validation on training set, with a model optimized for higher recall and negative predictive value.

In the multivariable ML model predicting all-cause mortality, CKD status was defined by the CKD-EPI equation, which emerged as the most influential renal predictor, followed by CKD status defined by the MDRD equation. Both inputs were included concurrently in the model as binary indicators, alongside other clinical predictors. The higher feature importance of CKD-EPI indicates a stronger independent contribution to mortality prediction compared with MDRD, rather than the selective inclusion of CKD-related deaths. Predictors that followed are age and comorbid conditions, which demonstrated stable importance rankings across models. For each predictor (EPI-CKD, MDRD CKD, age, diabetes duration in days, neuropathy, hypertension, congestive heart failure, myocardial infarction, nephropathy, gender, and retinopathy), values were randomly permuted 30 times in the test set, and the resulting decrease in model accuracy was recorded. Bar lengths represent the mean decrease in accuracy (permutation importance), and error bars indicate the standard deviation across permutations (Figure 4). As seen in this figure, the EPI formula seems to be the most important variable in the classification performance of mortality.

Figure 4.

Explainability SHAPLEY plots of the random forest classifier. (a) Feature importance ranking on the test set showing diabetes duration, age, and CKD-EPI as the most significant predictors of time-to-event death. (b) Waterfall SHAP plot for a selected example in a 65-year-old male with >4 years of diabetes duration and a predicted CKD risk using CKD-EPI.

To further interpret the model, feature permutation was applied. This technique quantifies the reduction in model performance when a predictor’s values are randomly shuffled, thus indicating its relative contribution to prediction accuracy. As illustrated in Figure 4, the CKD-EPI and DM duration/days, along with patient age, were identified as the top 3 predictors of mortality, substantially outperforming other clinical features such as comorbidities. The CKD-EPI formula demonstrated the highest importance score, underscoring its critical predictive value in mortality classification within this cohort.

Discussion

Our study evaluated the effectiveness of two equations in predicting mortality among Jordanian patients with DM, yielding promising results with CKD-EPI. The estimated eGFR categories were determined using both CKD-EPI and MDRD equations across Jordanian patients with DM. Although both equations exhibited agreement, MDRD tended to overestimate the higher stages of CKD. The CKD-EPI equation generally provides a more accurate estimation of the GFR, especially in patients with higher GFR values. This is consistent with the global findings that CKD-EPI tends to perform better in populations with normal or mildly reduced kidney function.^19,20

Commonly prescribed medications in the population with CKD can substantially influence kidney function and clinical outcomes. Renin–angiotensin system inhibitors such as angiotensin-converting enzyme inhibitors and angiotensin receptor blockers are crucial management that decrease proteinuria and CKD progression and improve cardiovascular and renal outcomes in randomized and observational studies. Those studies are strongly endorsed in contemporary KDIGO guidelines for CKD with proteinuria and cardiovascular risk reduction.²¹ On the other hand, non-steroidal anti-inflammatory drugs (NSAIDs) are categorized as nephrotoxins. Prolonged and short-course systemic NSAID use in older adults is associated with higher rates of acute kidney injury and hyperkalemia, with risk amplified in people with CKD, diabetes, cardiovascular disease and when combined with renin–angiotensin–aldosterone system blockade or diuretics. These medication effects interact with risk factors such as aging, sex, and ethnicity, comorbidities such as diabetes, hypertension, and cardiovascular disease, and unhealthy lifestyle factors such as smoking. Those are independent predictors of CKD progression. Consequently, significant confounders require meticulous management in observational outcome research.²²

Notwithstanding the emergence of contemporary GFR-estimating equations such as CKD-EPI, conventional formulas such as Cockcroft–Gault and MDRD continue to be extensively utilized for drug dosing, as the majority of historical pharmacokinetic studies and product labeling rely on creatinine clearance.²² No one equation is superior; CKD-EPI typically yields more precise GFR estimates, particularly at elevated GFR levels, and is currently endorsed by guideline organizations like KDIGO and professional pharmacy associations as the primary estimation method, whereas MDRD remains integral to drug development, regulatory labeling, and certain national formularies.²³ Regulatory agencies (KDIGO) emphasize that drug dosing should be based on absolute GFR (mL/min) and advocate using the equation specified in the drug profile. Direct measurement of clearance or therapeutic drug monitoring in particular, is required in high‑risk populations and for narrow‑therapeutic‑index drugs.²⁴

Due to the scarcity of nephrology care, the management of CKD is hindered among patients with DM. The integration of CKD coding within EHRs improves the identification of patients with persistently low eGFR. Tuttle et al.²⁰ found that 40% of their sample had CKD based on eGFR, 27% based on administrative coding, and 25% through both eGFR and administrative codes; patients with DM had the highest proportion of CKD. Inconsistent with our findings, CKD patients in their study were predominantly females aged 70 years and older compared with individuals at risk of CKD.²⁰ Tangri et al.⁸ proposed solutions for each CKD stage to be integrated into EHRs. Our study highlights the potential of EHRs as an optimal platform for implementing predictive models that integrate eGFR and ML to enhance CKD prevention, identification, and management.

Integrating precequations and ML models into EHRs for clinical use is increasingly advocated.³ RF analysis has identified CKD-EPI as superior in detecting CKD among both high and low-risk individuals compared with MDRD.^3,25 This finding aligns with Dhanwanth et al.²⁶ results, demonstrating the effectiveness of RF ensemble learning in mitigating CKD progression. However, there is limited understanding of advanced ML algorithms in nephrology research and eGFR. Unlike traditional statistical regression approaches, ML methods do not require prior knowledge of the data. Yet by leveraging historical data and advanced ML algorithms, hospitals can forecast future trends, enhance decision-making, and improve patients’ outcomes. Takkavatakarn et al.²⁷ compared RF models for CKD stage estimation, finding four ML models comparable and validated in predictive performance. Chuah et al.¹⁰ demonstrated that ML outperforms traditional statistical approaches in estimating GFR decline. Nevertheless, no research has yet used ML models to compare the predictive performance between the MDRD and CKD-EPI equations.^3,10 Clinical decisions regarding CKD detection and staging lie with nephrologists and clinicians; however, integrating predictive models by ML into EHRs could optimize outcomes.

Efforts to identify GFR estimation formulas that predict better healthcare outcomes have yielded inconsistent evidence. While MDRD achieved 51% accuracy, CKD-EPI reached 57% using traditional statistical analyses among diabetic patients. Minimal research using ML to estimate GFR exists, especially in the Jordanian population, and no studies have assessed the accuracy of GFR equations or CKD classification using ML methods. Additionally, the CKD-EPI equation classified fewer diabetic patients into higher eGFR categories compared with MDRD, potentially impacting risk categorization. While both equations estimate GFR, they differ in the variables considered and accuracy across populations. Despite the potential advantages of CKD-EPI, both equations are valuable for predicting outcomes and guiding treatment decisions in high-risk CKD patients, albeit imperfectly.²⁸

Measured GFR using exogenous substances is costly and time-consuming, whereas serum creatinine-based eGFR is cost-effective and widely used as a predicting factor for kidney function. Despite the variant validity and utility of different equations, including MDRD and CKD-EPI, inconsistencies persist in predicting outcomes. Clinicians and researchers must consider the limitations of equations such as MDRD and potentially favor CKD-EPI, especially in patients with preserved kidney function or higher GFR values.^13,29 In the present study, ethnicity was adjusted to reflect the predominantly Jordanian patient population, which may have contributed to the similar predictive performance of both equations for mortality. This finding aligns with Armani et al.,³⁰ who demonstrated that body surface area adjustment significantly improved agreement among six eGFR equations in a cohort of 60 Caucasian volunteers with varying severity of renal impairment.

Studies have shown varied performance between formulas in diagnosing CKD stages and predicting related health outcomes. Takkavatakarn et al.²⁷ equally identified MDRD and CKD-EPI as significant predictors of cardiovascular events among type II diabetes patients. In our study, mortality events were inversely related to GFR levels, with CKD-EPI demonstrating superior performance over MDRD and other covariates. ML models, like XGBoost, have identified eGFR and glucose as major predictors for ESRD, while traditional methods and ML models have shown similar performance in predicting ESRD among CKD stage 4 patients.^27,28 Our study contributes to the assessment of renal function in diabetic patients by integrating validated equations, traditional statistical methods, and ML approaches.

Limitation

The current study benefits from several strengths, including a substantial sample size, extensive follow-up duration, frequent serum creatinine sampling, and use of advanced statistical methodologies. However, it is important to acknowledge certain limitations. Our study was constrained by incomplete follow-up and insufficient clinical data on CKD. Furthermore, the study period encompassed the COVID-19 pandemic, introducing potential confounders because of its profound impact on diabetes care, thus influencing our findings across both pre- and post-pandemic data points.

Clinicians should choose the formula based on specific patient population characteristics and clinical scenarios. For CKD patients, both formulas can be useful, but CKD-EPI may offer better accuracy for the Jordanian patients with higher GFR. It is more accurate in estimating GFR in the Jordanian population with DM, including those with normal or near-normal kidney function. There may be a need for education and updates in clinical practice to ensure accurate GFR estimation and patient care. CKD-EPI is more accurate for the population diagnosed with DM and with normal or mild kidney function compared with the MDRD formula. Transitioning from MDRD to CKD-EPI may require updates to the clinical protocols and electronic health systems.

Conclusion

In conclusion, the CKD-EPI equation classifies fewer Jordanian patients with DM into higher eGFR categories. While CKD-EPI may perform better in certain aspects, both equations are useful tools for predicting outcomes and guiding treatment decisions in high-risk patients with DM. Moreover, integrating ML, traditional analysis methods, and validated equations is promising in enhancing the results of EHRs in preventing and detecting CKD among diabetic patients. The findings of this study have implications for researchers and clinicians in several aspects. Therapeutic interventions can be tailored according to the risk profile of CKD patients, allowing for the strategic planning of patient needs and the allocation of healthcare system resources.

Supplemental Material

sj-docx-1-taj-10.1177_27558428261447944 – Supplemental material for Comparative prognostic accuracy of MDRD and CKD-EPI equations in Jordanian diabetics: A machine learning-enhanced longitudinal study

Supplemental material, sj-docx-1-taj-10.1177_27558428261447944 for Comparative prognostic accuracy of MDRD and CKD-EPI equations in Jordanian diabetics: A machine learning-enhanced longitudinal study by Amani Anwar Khalil, Randa I. Farah, Mohammad Alrawashdeh and Salah Al-Zaiti in Sage Open Chronic Disease

Footnotes

Acknowledgements

The authors thank the Deanship of Scientific Research at the University of Jordan for funding this research, and many thanks go to the EHRs (HAKEEM) for their cooperation in providing the research dataset.

ORCID iD

Amani Anwar Khalil

Ethical considerations

The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the Ministry of Health in Jordan and the Deanship of Scientific Research at the University of Jordan (12/2022-2023) on April, 22, 2023.

Consent to participate

In accordance with institutional review board (IRB) regulations, the requirement for informed consent was waived due to the retrospective nature of the study.

Consent for publication

The study was conducted using retrospectively collected data. All patient information was fully anonymized prior to analysis, and no identifiable personal data are included in this manuscript.

Author contributions

Amani Anwar Khalil: Conceptualization; Data curation; Funding acquisition; Writing – original draft.

Randa I. Farah: Investigation; Methodology; Writing – review & editing.

Mohammad Alrawashdeh: Formal analysis; Writing – review & editing.

Salah Al-Zaiti: Data curation; Validation; Visualization; Writing – review & editing.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: the research was funded by the Deanship of Scientific Research at the University of Jordan.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The datasets generated and/or analyzed during the current study are not publicly available due to institutional and ethical restrictions related to patient confidentiality.

Supplemental material

Supplemental material for this article is available online.

References

World Health Organization. Global report on diabetes. France, file:///D:/sabatical/WHO DM report.pdf (2016, accessed 13 July 2022).

Momani

Haddad

Barham

, et al. Current practice in the management of type 2 diabetes mellitus in Jordan: results of the Jordanian Diabetes Management Practices Study. Int J Diabetes Res 2020; 9(1): 12–22.

Chen

Kantagowit

Nopsopon

, et al. Prediction and diagnosis of chronic kidney disease development and progression using machine-learning: protocol for a systematic review and meta-analysis of reporting standards and model performance. PLoS One 2023; 18(2): e0278729.

Khalil

Abed

Ahmad

, et al. Under-diagnosed chronic kidney disease in Jordanian adults: Prevalence and correlates. J Ren Care 2018; 44(1): 12–18.

Alramly

Darawad

Khalil

. Slowing the progression of chronic kidney disease: a comparison between predialysis and dialysis Jordanian patients. Ren Fail 2013; 35(10): 1348–1352.

Palacio-Lacambra

M-E

Comas-Reixach

Blanco-Grau

, et al. Comparison of the Cockcroft—Gault, MDRD and CKD—EPI equations for estimating ganciclovir clearance. Br J Clin Pharmacol 2018; 84(9): 2120–2128.

Schwandt

Denkinger

Fasching

, et al. Comparison of MDRD, CKD-EPI, and Cockcroft-Gault equations in relation to measured glomerular filtration rate among a large cohort with diabetes. J Diabetes Complications 2017; 31(9): 1376–1383.

Tangri

Ferguson

. Practical utilization of prediction equations in chronic kidney disease. Blood Purif 2023; 51(Suppl. 1): 32–38.

Tangri

Grams

Levey

, et al. Multinational assessment of accuracy of equations for predicting risk of kidney failure: a meta-analysis. JAMA 2016; 315(2): 164–174.

10.

Chuah

Walters

Christiadi

, et al. Machine learning improves upon clinicians’ prediction of end stage kidney disease. Front Med (Lausanne) 2022; 9: 837232.

11.

Zheng

Xue

B-W

Guo

A-H

, et al. Patient delay in chronic kidney disease: a qualitative study. Medicine (Baltimore) 2023; 102(48): e36428.

12.

Ndapkwi

Zeuko

Mpey

, et al. Prevalence, risk factors, and awareness of chronic kidney disease among diabetes mellitus and hypertensive individuals at the Buea Regional Hospital, Cameroon. J Biosci Med 2023; 95: 260–277.

13.

Ajlouni

Batieha

Jaddou

, et al. Time trends in diabetes mellitus in Jordan between 1994 and 2017. Diabet Med 2019; 36(9): 1176–1182.

14.

Hammoudeh

Tabbalat

Alhaddad

, et al. Short- and long-term outcomes in Middle Eastern diabetic patients after percutaneous coronary intervention: results from The First Jordanian PCI Registry (JoPCR1). Diabetol Int 2017; 8(1): 30–38.

15.

Dagenais

Radi

Burnham

, et al. Comparing five equations to estimate glomerular filtration rate or creatinine clearance and assign individuals to KDIGO categories across the full age spectrum using real-world data. Clin Transl Sci 2024; 17(2): e13726.

16.

The World Bank. World Bank country and lending groups [Internet]. Washington, DC: The World Bank, https://datahelpdesk.worldbank.org/knowledgebase/articles/906519-world-bank-country-and-lending-groups (2026, accessed 13 May 2025).

17.

Kovesdy

. Epidemiology of chronic kidney disease: an update 2022. Kidney Int Suppl 2022; 12(1): 7–11.

18.

De Souza dos Santos

Aparecida de Oliveira Duarte

Maria Trevisan Zanetta

. Prevalence and incidence of decreased glomerular filtration rate and its variation over 6 years: cohort study SABE 2010–2016. PLoS One 2024; 19(1): e0294660.

19.

Alicic

Rooney

Tuttle

. Diabetic kidney disease: challenges, progress, and possibilities. Clin J Am Soc Nephrol 2017; 12(12): 2032–2045.

20.

Tuttle

Alicic

Duru

, et al. Clinical characteristics of and risk factors for chronic kidney disease among adults and children: an analysis of the CURE-CKD Registry. JAMA Netw Open 2019; 2(12): e1918169.

21.

Covic

Voroneanu

Stefan

, et al. Comparative renal effects of angiotensin receptor neprilysin inhibitors and ACEi/ARB: a systematic review and meta-analysis. Clin Kidney J 2025; 18(8): sfaf224.

22.

Mirkov

Scuderi

Lloyd

, et al. Estimation of kidney function for medication dosing in adult patients with chronic kidney disease: a practice update. J Pharm Pract Res 2024; 54(1): 94–106.

23.

Donker

Bet

Nurmohamed

, et al. Estimation of glomerular filtration rate for drug dosing in patients with very high or low body mass index. Clin Transl Sci 2022; 15(9): 2206–2217.

24.

Levey

Inker

. Assessment of glomerular filtration rate in health and disease: a state of the art review. Clin Pharmacol Ther 2017; 102(3): 405–419.

25.

Ouyang

Chen

, et al. Self‑management program for patients with chronic kidney disease (SMP‑CKD) in Southern China: protocol for an ambispective cohort study. BMC Nephrol 2022; 23(1): 93.

26.

Dhanwanth

Vivek

Abirami

, et al. Forecasting chronic kidney disease using ensemble machine learning technique. IJRITCC 2023; 11: 336–344.

27.

Takkavatakarn

Cheng

, et al. Machine learning models to predict end-stage kidney disease in chronic kidney disease stage. BMC Nephrol 2023; 24(1): 376.

28.

Wang

Bowe

Cui

, et al. A deep learning approach for the estimation of glomerular filtration rate. IEEE Trans Nanobiosci 2022; 21(4): 560–569.

29.

Goro

Wolide

Dibaba

, et al. Patient awareness, prevalence, and risk factors of chronic kidney disease among diabetes mellitus and hypertensive patients at Jimma University Medical Center, Ethiopia. Biomed Res Int 2019; 2019: 2383508.

30.

Armani

Stingl

Forst

, et al. Retrospective analysis of glomerular filtration rate estimated using different equations with and without adjustment to individual body surface area. Med Drug Discov 2025; 27(June): 2–6.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.11 MB

0.00 MB