Development and validation of a risk prediction model for chronic kidney disease among adult hypertensive patients having follow-up at University of Gondar Comprehensive Specialised Hospital, Ethiopia: a retrospective cohort study

Abstract

Objective

Chronic kidney disease (CKD) arises due to uncontrolled hypertension (HTN). HTN significantly increases the risk of complications in vital organs, mainly the kidneys. If hypertensive individuals receive early intervention, the majority of these complications and deaths from CKD can be avoided. Having a clinically applicable tool to predict the future risk of those complications can prevent early disability and premature mortality. However, to this day, there is a lack of a validated risk prediction model specifically designed for CKD of hypertensive patients in Ethiopia. We aimed to develop a risk prediction model for CKD among hypertensive patients at the University of Gondar Comprehensive Specialised Hospital (UoGCSH), Ethiopia.

Study design

A retrospective follow-up study was conducted from 1 January 2012 to 30 December 2021. The Least Absolute Shrinkage and Selection Operator regression methods were used to select predictors. The performance of the models was assessed using the Area Under the Curve and calibration plots. The internal validity of the model was evaluated using bootstrapping methods, and the model was presented as a nomogram. Decision curve analysis was conducted to assess the net benefit of the prediction model in clinical and public health contexts.

Setting

Data from patients’ medical records were collected via the Kobo Toolbox in the UoGCSH.

Participant

We followed a total of 1120 Patients diagnosed with HTN.

Results

The incidence of CKD among adult hypertensive patients was 19.82% (95% CI 17.59% to 22.26%). In the multivariable logistic regression analysis, age, residency, baseline blood pressure status, type of HTN, family history of HTN, baseline serum creatinine levels, proteinuria at baseline and dyslipidaemia were identified as statistically significant predictors of CKD. The nomogram demonstrated a discriminatory power of 91.98% (95% CI 90.09% to 93.88%) and a calibration p value of 0.327. The sensitivity and specificity of the prediction model were 80.63% (95% CI 74.81% to 85.61%) and 87.97% (95% CI 85.66% to 90.03%), respectively. The developed nomogram has a greater net benefit than using the treat-all or treat-none strategies when the threshold probability of the patient is increased.

Conclusion

The nomogram demonstrated excellent discrimination and calibration in identifying hypertensive patients at high risk of CKD. This predictive model offers clinicians a valuable tool for early identification of high-risk individuals, enabling timely interventions, personalised counselling and optimised management through close monitoring to prevent disease progression.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• First, this study benefited from a relatively large sample size and an extended follow-up period.

• Second, to our knowledge, this is the first nomogram developed for the prediction of chronic kidney disease in the country.

• A notable limitation is that the reliance on secondary data precluded the inclusion of important predictors, such as participants’ body mass index, educational status, marital status, salt intake and socioeconomic status.

• The study was conducted at a single centre in Ethiopia, limiting its generalisability to other settings.

Introduction

Statement of the problem

Chronic kidney disease (CKD) is an increasingly critical global public health issue that can lead to renal failure, cardiovascular disease and premature death. It is characterised by reduced kidney function, usually indicated by an estimated glomerular filtration rate (eGFR) of <60 mL/min/1.73 m²¹ or by evidence of kidney damage, such as proteinuria, that persists for at least 3 months.2,4

More than 10% of the global population is affected by CKD, leading to millions of deaths each year due to the inability to afford necessary treatments.⁵ The global prevalence of CKD ranges from 8% to 16%, with nearly 500 million people worldwide believed to be living with this condition; the majority (78% or 387.5 million) reside in low-income to middle-income countries.⁶ In 2017, CKD was the 12th leading cause of death worldwide.⁷ Incidence rates in low-income to middle-income countries may be up to four times higher than in developed countries. The prevalence of CKD among the general populations of India, Korea, Pakistan and Southern China is 17.2%, 13.7%, 12.5% and 9.4%, respectively.8,11 End-stage renal disease (ESRD) typically occurs in individuals aged 20 to 50 in sub-Saharan Africa, manifesting 20 years earlier than in other ethnic groups in Western countries.^{6 12}

Hypertension (HTN) is a leading cause of CKD, resulting from damage to the renal vasculature due to elevated blood pressure.13,15 The relationship between HTN and CKD is cyclical; HTN not only contributes to CKD but also results from renal impairment. In the United States, HTN is the second leading cause of ESRD, with accelerated CKD progression observed in hypertensive patients.16,18 Approximately 6% of individuals with essential HTN also have CKD, placing them at increased risk for progression to ESRD. The Seventh Report of the Joint National Committee and the National Kidney Foundation-Dialysis Outcomes Quality Initiative classifies CKD as a high-risk condition, recommending blood pressure control with a target of <130/80 mm Hg.¹⁹

In the United States, 17.3% of people with prehypertension and 22.0% of those with undiagnosed HTN had CKD.²⁰ A study in Denver showed that 12.1% of hypertensive patients developed CKD over 7 years.²¹

In Asian countries, the incidence of CKD among HTN patients ranged from 5% to 20%, with China, India and Taiwan having the highest rates.²² In Spain, 40.6% of essential hypertensive patients developed CKD,²³ and a South Korean study found that 8.2% of hypertensive patients developed CKD over ten years, with an incidence rate of 9.67 per 1000 person-years.²⁴

A systematic review revealed that the prevalence of CKD in Africa ranges from 2% to 41%, with HTN patients experiencing a prevalence of 13% to 51%.¹² In sub-Saharan Africa, the prevalence is estimated at 32.3%,²⁵ and 98 studies conducted in Africa at 15.8%.²⁶ Other studies have reported CKD prevalence among hypertensive patients at 7.7% in Egypt, 26.3% in Ghana and 50.5% in Nigeria.^{27 28} In Ethiopia, prevalence rates range from 17.6% to 26%.29,32

Key sociodemographic factors associated with CKD include older age^{33 34} or increasing age,^{23 29 35} female sex^{25 35} (except for Ghana, where male sex was a predictor^{28 36}), and low educational status.³⁶ Clinical predictors include systolic and diastolic blood pressure, serum creatinine levels, the presence of comorbid diabetes mellitus,^{21 23 25 29 34} uncontrolled blood pressure and fasting blood sugar.^{29 32} Other significant clinical factors include the duration and stage of HTN, proteinuria, dyslipidaemia²⁹ and elevated serum creatinine. Specific regional studies also highlight variations: in Ethiopia,³¹ predictors such as Stage II HTN,³⁷ proteinuria and dyslipidaemia were significant, while in Japan and Hong Kong,³⁸ Stage II HTN and increased systolic blood pressure were prominent.²⁸ In South India, body mass index (BMI), diabetes mellitus and HTN severity were critical factors,³⁵ while in South Korea, uncontrolled blood pressure was highlighted.²⁴ Finally, in China, additional predictors such as stroke, low-density lipoprotein levels and duration of HTN were also associated with CKD development among hypertensive individuals.³³

Ethiopia is committed to achieving the Sustainable Development Goals from 2016 to 2030, including reducing premature deaths from non-communicable diseases by one-third.³⁹ However, HTN remains the leading cause of CKD, leading to end-stage kidney disease that requires renal replacement therapy, with significant social and economic impacts.⁴⁰ The burden of CKD in Ethiopia is rising due to increasing non-communicable diseases associated with sedentary lifestyles.⁴¹ Despite this, evidence for predicting CKD is limited. This study aims to identify high-risk hypertensive patients using a prognostic risk score model. The developed nomogram is designed for easy clinical use, incorporating accessible predictors that do not require advanced clinical assessments.

Developing a prognostic risk score for CKD based on demographic and clinical factors is essential in the Ethiopian context. Researchers advocate for early identification and management of high-risk patients in low-resource settings through prediction models. Providing alternative clinical prediction tools that use easily accessible patient information will help healthcare professionals save lives and reduce costs for hypertensive patients. Regular check-ups can incorporate this tool to identify and manage high-risk patients effectively, thus preventing the onset of CKD. For policymakers, this approach aids in planning and delivering effective healthcare services. Overall, the new risk score developed in this study could help reduce illness and deaths related to CKD at local, regional and national levels.

Predicting the risk of CKD is vital for quickly identifying high-risk patients and initiating necessary treatments for complications related to high blood pressure, ultimately reducing morbidity and mortality. Therefore, this study aims to develop and validate a risk prediction model for predicting CKD among hypertensive patients.

Methods

Study design and period

A retrospective follow-up study was conducted at the University of Gondar Comprehensive Specialised Hospital (UoGCSH) on patients diagnosed with high blood pressure from 1 January 2012 to 30 December 2021. The follow-up observation period lasted for 10 years. Data were extracted from 11 March 2022 to 11 May 2022.

Study area

The study was carried out at the UoGCSH, which is situated in the town of Gondar, 743 km northwest of Addis Ababa, the capital of Ethiopia, and 178 km from Bahir Dar, Amhara Regional State. The hospital serves the majority of the community in Amhara Regional State and its surrounding areas. It is a teaching facility dedicated to community service, research and education. It provides service for over 5 to 7 million people. With almost 500 inpatient beds, the facility offers a variety of disciplines, including internal medicine, psychiatry, paediatrics, surgery, gynaecology/obstetrics, HIV care and an outpatient clinic. The outpatient centre comprises a CT scan room, pharmacy and departments for MRIs and ultrasounds.

The hospital also offers dialysis services. Services in the chronic follow-up care unit are provided with the support of nurses, general practitioners, senior specialists and residents. From January 2012 to December 2021, about 4925 adult HTN patients came to the department for services linked to chronic care.

Population

A list of hypertensive patients is kept in the registration book of the UoGCSH. First, the medical records of patients who underwent follow-up and were recorded from January 2012 to December 2021 in the hospital were isolated. Those who were under 18 were excluded based on their medical record numbers. Ten years of registered hypertensive patients were included in this study. Patient charts for those who developed CKD before HTN were excluded. Next, 1120 adult hypertensive patient charts were selected using a simple random sampling technique after providing continuous numbers for each chart number. These patients had follow-up visits from 1 January 2012 to 31 December 2021, and were followed up for at least 3 months at the UoGCSH. Patients’ charts were chosen based on their medical record numbers. Six charts were absent and were replaced by other randomly selected charts. Lastly, the selected medical charts were followed until the outcome was developed.

Clinical and sociodemographic characteristics and definition

The following baseline information was collected: age, gender, residence, comorbid with diabetes mellitus, systolic blood pressure, diastolic blood pressure, type of HTN, family history of HTN, stage of HTN, blood pressure status (was measured as uncontrolled blood pressure if there is three consecutive follow-up BP ≥140/90 mm Hg without diabetes mellitus and ≥130/80 mm Hg with diabetes mellitus⁴²), serum creatinine level, fasting blood glucose level, Dyslipidaemia (defined as TC ≥200 mg/dL, or TG ≥150 mg/dL, or LDL-C ≥130 mg/dL, or HDL-C <50,⁴³ proteinuria (according to the laboratory result, if a patient had plus one and above at baseline, we classify as having proteinuria⁴⁴), and alcohol use. Finally, CKD was defined and identified when physicians diagnosed the patient as having CKD.³¹ The physicians used an eGFR, the presence of proteinuria for more than 3 months and ultrasound evidence of kidney size decline to diagnose CKD.

Data collection procedure and quality control

All relevant data were collected retrospectively from patient charts. The English version extraction checklist was developed from different literature after being customised according to the available variables in the patient’s chart. The proposed checklist was changed to an electronic data collection tool called KoBo Toolbox and deployed after installing the KoBo Collect mobile application on each data collector’s smartphone device. The data-gathering method involved four BSc nurses as data collectors and one MSc nurse as a supervisor. Nurses in the chronic follow-up care unit extracted the required data from patient charts under supervision in the hospital medical records room.

Two clinical experts with specialised knowledge in the relevant medical field and an experienced researcher skilled in data analysis and methodology rigorously reviewed and validated the extraction tool. They assessed it comprehensively for content accuracy, ensuring all critical information was captured, and evaluated clarity to guarantee ease of use and consistency during the data extraction to ensure they accurately reflected the study’s objectives.

A pretest was conducted involving 40 participants, representing 5% of the total sample size, to confirm the presence of relevant variables in the patient records. Additionally, it ensured that the data collectors and supervisors were adequately trained and competent to oversee the data collection process.

As a result, the number of unrecorded variables like BMI, smoking, physical activity, salt intake and socioeconomic status in the data extraction tool was minimised. One day of training was given to data collectors and supervisors. The training covered how to use the KoBo Collect digital extraction tool, access records, approach each item in the patient records and handle data. Both the principal investigator and supervisor closely supervised the entire data collection process. During the data extraction process, the assigned supervisor and principal investigator closely monitored the work, ensuring the collected data was complete and accurate. This quality checking was done daily after data collection and corrections were made.

Patient and public involvement

Patients and/or the public were not involved in this research’s design, conduct, reporting or dissemination plans.

Data processing and analysis

The collected data using KoBo Collect was exported to MS Excel for data management. Finally, the analysis was done using STATA V.17 statistical software.

Patients’ sociodemographic characteristics were described using descriptive analysis of central tendency (mean), dispersion (SD) and frequency distribution of categorical data. Variables were used as they were registered on the chart. However, when reasonable, categorizations were made based on clinical relevance and previous literature for the sake of appropriate prediction and easy applicability; such categorizations were made.

During model development, categorical variables were checked for χ² assumptions. Then, the selected variables were entered into the least absolute shrinkage and selection operator (LASSO) regression model. Then, variables with non-zero coefficients from the LASSO model were in the multivariable logistic regression. Therefore, the multivariable logistic regression model was fitted. Significant variables with a p value <0.1 were incorporated into the reduced model, and model comparison was done using the log-likelihood ratio test. Finally, the predictors in the reduced model, a CKD risk prediction tool (nomogram), were developed.

Overall model performance was evaluated by the difference between the predicted outcome and the actual outcome using two characteristics of performance measures: calibration and discrimination.⁴⁵ The nomogram’s performance was evaluated by the area under the curve (AUC) of the receiver operating characteristics to assess the model’s discriminating ability and the calibration plot to find agreement between the observed and predicted probability of developing CKD.⁴⁶ The model’s performance was also evaluated by computing its accuracy, sensitivity, specificity and predictive values.⁴⁷ Youden’s index approach was used to determine the probability cut-off point for risk classification as high or low.⁴⁸ A customised risk stratification table was created using the determined cut-off point to classify each hypertensive patient as either high-risk or low-risk and enable preventative measures based on the assessed risk of developing CKD.⁴⁹ The model’s internal validity was assessed using the bootstrapping method. The bootstrapping method was applied to internal validation, and it is an ideal method for smaller sample sizes.⁵⁰ Bootstrapping yields a reliable estimate with minimal bias compared with other internal validation techniques like the split sample.^{51 52} The performance before validation and the performance after validation were calculated. The transparent reporting of a multivariable prediction model for an individual prognostic or diagnostic checklist was used to report the study’s findings.⁵³ The decision curve analysis (DCA) was assessed for the net benefit of using the developed model by clinicians in clinical practice.⁵⁴

Results

Sociodemographic characteristics of adult hypertensive patients

A total of 1120 patient medical records were reviewed. Among these records, over 60.0% (676) were female, and the majority, 62.77% (703), resided in urban areas. At the beginning of the follow-up period, the participants’ ages ranged from a minimum of 22 years to a maximum of 86 years, with a mean age of 54.98 years (SD: 12.05 years).

Baseline clinical characteristics of adult hypertensive patients

The systolic blood pressure readings ranged from a minimum of 140 mm Hg to a maximum of 251 mm Hg, with a mean of 160.83 mm Hg (SD: 18.15 mm Hg). For diastolic blood pressure, the minimum was 78 mm Hg and the maximum was 180 mm Hg, resulting in a mean of 93.65 mm Hg (SD: 11.15 mm Hg).

The average duration of follow-up for participants with HTN was 3.85 years (SD: 2.84 years), with the follow-up duration ranging from a minimum of 0.252 years to a maximum of 9.95 years.

Fasting blood glucose levels varied from a minimum of 78 mg/dL to a maximum of 343 mg/dL, with an average fasting blood glucose of 111.71 mg/dL (SD: 34.05 mg/dL).

Among the participants, 876 (78.21%) had primary HTN, and nearly half (45.27%) began their follow-up with stage I HTN. More than half (52.23%) of the individuals had a family history of HTN, while a smaller percentage (27.23%) of hypertensive patients had uncontrolled blood pressure (table 1).

Table 1. Baseline clinical characteristics of the adult hypertensive patients who have follow-up at University of Gondar comprehensive specialised hospital from 1 January 2012 to 30 December 2021 (n=1120).

Categories	Per cent	Chronic kidney disease		Frequency	Variables
		Yes	No
Type of hypertension	Primary	153	723	876	78.21
	Secondary	69	175	244	21.79
Stage of hypertension	Stage I	71	507	502	45.27
	Stage II	103	327	399	38.39
	HTN Crisis	48	183	219	16.34
Alcohol use	Yes	104	318	422	37.68
	No	118	580	698	62.32
Presence of proteinuria	Yes	171	135	306	27.32
	No	51	763	814	72.68
Serum creatinine level	High	172	134	306	27.32
	Normal	50	764	814	72.68
Comorbidity with diabetes mellitus	Yes	72	253	325	29.02
	No	150	645	795	70.98
Presence of dyslipidaemia	Yes	69	239	308	27.50
	No	153	659	812	72.50
Blood pressure status	Uncontrolled	97	305	305	27.23
	Controlled	125	815	815	72.77

HTN, hypertension.

Incidence of chronic kidney disease among hypertensive patients having follow-up at UoGCSH

The incidence proportion of CKD among adult hypertensive patients was 19.82% (95% CI 17.59% to 22.26%).

Development and validation of a chronic kidney disease prediction model

A total of 15 predictor variables were included in the LASSO regression model, and non-zero coefficients were identified at a tuning parameter (lambda) of 0.001611 with a cross-validated mean deviance of 0.5613315 and an out-of-sample deviance ratio of 0.4363 at the 54th iteration. All variables included in the LASSO had non-zero coefficients and were thus included in the multivariable logistic regression model. The multivariable logistic regression model was fitted with all variables of non-zero coefficients. Model reduction was conducted by reducing insignificant variables (p>0.1) using the likelihood ratio test to achieve model parsimony. Seven variables like gender (p=0.255), stage of HTN (p=0.574), alcohol use (p=0.232), comorbidity with diabetes mellitus (p=0.543), systolic blood pressure (p=0.598), diastolic blood pressure (p=0.576) and fasting blood glucose level (p=0.497) were removed. The full model (model with nine predictors) and the reduced model (model with seven predictors) were compared using the likelihood ratio test, which implies that no statistically significant difference was observed between the two models (likelihood ratio χ²=10.46 and p=0.2322). Finally, eight predictors were retained and used for nomogram development (table 2).

Table 2. Multivariable logistic regression analysis and model reduction using potential predictors of chronic kidney disease among adult hypertensive patients who have follow-up at University of Gondar comprehensive specialised hospital from 1 January 2012 to 30 December 2021 (n=1120).

Variable	Category	β-coefficient (95% CI)	P value	β-Coefficient	P value
Age		0.27 (0.19 to 0.73)	0.010	0.02 (0.01 to 0.04)	0.009
Gender	Male	0.28 (−0.17 to 0.73)	0.255
	Female	1
Residency	Rural	1.49 (1.05 to 1.94)	0.000	1.52 (1.09 to 1.96)	0.000
	Urban	1		1
Family history of hypertension	Yes	1.26 (0.82 to 1.70)	0.000	1.24 (0.81 to 1.68)	0.000
	No	1		1
Type of hypertension	Primary	1		1
	Secondary	0.72 (0.19 to 1.24)	0.007	0.53 (0.05 to 1.01)	0.029
Stage of hypertension	Stage I	1
	Stage II	0.24 (−0.34 to 0.83)	0.411
	Hypertensive crisis	0.19 (−0.82 to 1.21)	0.709
Blood pressure status	Uncontrolled	0.96 (0.52 to 1.43)	0.000	1.08 (0.63 to 1.53)	0.000
	Controlled	1		1
Systolic blood pressure		0.01 (−0.01 to 0.03)	0.598
Diastolic blood pressure		0.01 (−0.013 to 0.025)	0.576
Comorbidity with diabetes mellitus	Yes	−0.19 (−0.84 to 0.44)	0.543
	No	1
Presence of proteinuria	Yes	1.45 (0.92 to 1.98)	0.000	1.5 (1.04 to 2.04)	0.000
	No	1			1
Serum creatinine level	High	2.19 (1.65 to 2.74)	0.000	2.16 (1.64 to 2.68)	0.000
	Normal	1		1
Dyslipidaemia	Yes	1.45 (0.10 to 1.04)	0.017	0.54 (0.08 to 0.99)	0.022
	No	1		1
Fasting blood glucose level		−0.003 (−0.011 to 0.005)	0.497
Alcohol use	Yes	0.27 (−0.17 to 0.71)	0.232
	No
Intercept		−7.92 (−11.32 to 4.51)	0.000	−6.51 (−7.75 to 5.28)	0.000

Based on the risk prediction model, the probability of developing CKD among adult hypertensive patients was predicted using the following regression formula, which is formulated with statistically significant prognostic predictors.

Probability of CKD (t)=${\beta }_0$+Age+residency+family history of HTN+Type of HTN+baseline BP status+baseline serum creatinine+baseline proteinuria+Dyslipidaemia.

Probability of CKD (t)=−6.513+0.025 age+1.524 rural+1.244 yes for family history of HTN+0.534 secondary type of HTN+1.079 uncontrolled HTN+2.159 high baseline serum creatinine+1.544 yes for proteinuria+0.536 yes for dyslipidaemia.

Nomogram for predicting the risk of chronic kidney disease among adult hypertensive patients

The prediction model was constructed by multivariable logistic regression model coefficients based on the eight identified predictors (age, residence, family history of HTN, type of HTN, blood pressure status, baseline serum creatinine, baseline proteinuria and dyslipidaemia) that can be easily identified to assist healthcare providers in discriminating patients with a high risk of CKD, enabling them to implement preventive measures as early as possible. To make it simple for clinicians to determine the predicted probability of CKD in adult hypertensive patients, the prediction model was developed, internally validated and then transformed into a nomogram. The nomogram visualises the model in a user-friendly manner (figure 1).

Figure 1. The nomogram for predicting the risk of chronic kidney disease (CKD) in adult hypertensive patients at UoGCSH. An individual with a total score of 22 will correspond to a 0.1572 probability of developing CKD; hypertensive patients with a ≥0.1572 probability will be classified as high-risk, whereas hypertensive patients with a <0.1572 probability will be classified as low risk during the application of the nomogram.

Performance of the nomogram

The contribution of each predictor to the overall AUC was presented (figure 2). The AUC of the nomogram for individualised CKD risk prediction was 91.99% (95% CI 90.07% to 93.91%) (figure 3).

Figure 2. The receiver operating characteristic curve shows the performance of each predictor. The diagonal line represents a model that discriminates by chance (area under the curve=50%); the x-axis shows the proportion of individuals without chronic kidney disease (CKD) who were incorrectly classified as having CKD (false-positive rate), and the y-axis shows the proportion of individuals with CKD who were correctly classified as having CKD (true positive rate). HTN, hypertension.

Figure 3. The receiver operating characteristic curve shows the performance of the original and internally validated model. The diagonal line represents a model that discriminates by chance (AUC=50%); the x-axis shows the proportion of individuals without chronic kidney disease (CKD) who were incorrectly classified as having CKD (false-positive rate), and the y-axis shows the proportion of individuals with CKD who were correctly classified as having CKD (true positive rate). AUC, area under the curve.

Using the coefficients (β), the predicted risk cut-off point identified by Youden’s index (max J=0.6856) was a probability of >0.1572. The model had a sensitivity of 86.04% (95% CI 80.77% to 90.31%), specificity of 82.52% (95% CI 79.87% to 84.95%), positive predictive value of 54.88% (95% CI 49.49% to 60.19%), negative predictive value of 95.98% (95% CI 94.35% to 97.26%) and accuracy of 83.21% (95% CI 80.89% to 85.36%). The calibration plot had a p value of 0.367, indicating no significant difference between the observed probability of developing CKD and the expected probability of developing CKD (figure 4A).

Figure 4. The calibration plot showing the agreement between observed (y-axis) and expected (x-axis) probabilities of developing chronic kidney disease at 95% CIs for the original (A) and internally validated models (B).

Chronic kidney disease risk stratification table

The risk stratification table was constructed using Youden’s index (max J=0.6856), and the corresponding probability cut-off point with this index was 0.1572. The risk was dichotomised into low (<0.1572) and high-risk (≥0.1572) groups. Of the participants, 348 (31.07%) were in the high-risk group. Of all CKD patients, the majority (54.89%) were from the high-risk category. The prevalence of CKD was 4.02% in low-risk groups (table 3).

Table 3. Risk stratification table based on the probability of developing chronic kidney disease identified by the nomogram.

Risk category (probability)	Frequency	Prevalence of CKD
Low risk (p<0.1572)	772 (68.93%)	31 (4.02%)
High risk (p≥0.1572)	348 (31.07%)	191 (54.89%)
Total	1120 (100%)	222 (19.82%)

CKD, chronic kidney disease.

Internal validation

The developed nomogram was internally validated by bootstrapping using 1000 bootstrap samples with replacement to determine the degree of overfitting. The AUC of the internally validated model was 91.98% (95% CI 90.09% to 93.88%) (figure 3). In addition, the model’s sensitivity was 80.63% (95% CI 74.81% to 85.61%), specificity was 87.97% (95% CI 85.66% to 90.03%), positive predictive value was 62.37% (95% CI 56.49% to 67.99%), negative predictive value was 94.84% (95% CI 93.11% to 96.24%) and accuracy was 86.52% (95% CI 84.38% to 88.46%) at the 0.2463 probability cut-off point identified by a Youden’s index of 0.6860.

The p value of the calibration plot was 0.327, which indicated the presence of agreement between the observed probability of developing CKD and the predicted probability of developing CKD among hypertensive patients (figure 4B). The optimism coefficient was 0.001.

Decision curve

The following DCA plot demonstrates that, compared with intervening on all or none of the hypertensive patients, the net benefit of using the model to carry out a specific intervention to address CKD in hypertensive patients is higher. The decision curve showed that using the developed risk prediction nomogram adds more benefit than the intervention for all or none of the patients (figure 5).

Figure 5. Decision curve analysis for the nomogram plotting net benefit of the model against threshold probability at UoGCSH, 2012–2021. CKD, chronic kidney disease.

Discussion

This study aimed to develop a risk prediction model for CKD among hypertensive patients. Hypertensive patients are at a higher risk of developing CKD than the general population because HTN is the second leading cause, following diabetes mellitus.⁴⁴ Developing a risk prediction model for this high-risk population plays a key role in the prevention strategy and early detection of CKD and its complications. It will contribute to the global target to reduce premature mortality from non-communicable diseases by a third by 2030.^{55 56}

The risk prediction model was developed using multivariable logistic regression with eight predictor variables selected by LASSO regression. Internal validation was conducted using bootstrapping. The resulting nomogram is a practical tool that helps healthcare professionals identify individuals at high risk for CKD during routine follow-ups.⁵⁷

The overall incidence proportion of CKD among hypertensive patients at the end of follow-up was 19.82%. This finding aligns with prevalence studies in the Tigray region (22.1%)²⁹ and Gondar (17.6%).³¹ It is lower than rates reported in sub-Saharan Africa (33.3%),²⁵ Egypt (33%),⁵⁸ Ambo (20.5%),³⁰ Jimma (26%)³² and Malaysia (30.9%),⁵⁹ but higher than in Colorado (12.1%).²¹ Multiple factors can influence the prevalence of CKD, including sample size, study context, duration and participant demographics.

Keeping other variables constant/controlled, each additional year of age increases the mean risk of CKD in adult hypertensive patients by 2%. This is supported by studies in Addis Ababa,⁶⁰ Spain,⁶¹ Peru,⁶² Kenya,⁶³ Tigray in Ethiopia,²⁹ Ambo in Ethiopia³⁰ and in the Savanna zone of Cameroon.²⁵ These changes may be attributed to physiological factors associated with ageing, such as decreased kidney mass, reduced nephron numbers and altered renal blood flow, which contribute to the onset and progression of CKD. Additionally, ageing significantly affects renal haemodynamics by disrupting renal autoregulation and blood pressure mechanisms, leading to increased glomerular pressure, proteinuria and a higher risk of kidney damage.⁶⁴

The odds of CKD among hypertensive patients living in rural areas were 4.57 times higher than those in urban areas. Studies from Nigeria indicated that half of rural hypertensive patients developed CKD.²⁷ This may be due to the significant impact of non-communicable diseases on rural residents with limited access to specialised healthcare. While the burden of HTN is greater in urban areas, low socioeconomic status and lower health literacy in rural areas can delay follow-up care.

In hypertensive patients with a family history of HTN, the odds of developing CKD were 3.46 times higher than in those without such a history. This finding is supported by a study in Sri Lanka⁶⁵ and Croatia.⁶⁶ A family history of HTN is common in primary HTN and increases the risk of earlier onset. Additionally, early-onset HTN significantly raises the risk of CKD by prolonging the kidneys’ exposure to elevated blood pressure.

The odds of developing CKD among hypertensive patients with secondary HTN were 1.69 times higher than in those with primary HTN. This may be because high blood pressure in secondary HTN can damage the delicate blood vessels and filters in the kidneys. Over time, this damage reduces the kidneys’ ability to filter waste effectively, contributing to CKD.⁶⁷ Additionally, secondary HTN is often more severe and harder to control, placing greater strain on the kidneys and accelerating the decline in kidney function.

We found that the odds of developing CKD in adult hypertensive patients with uncontrolled blood pressure at baseline were nearly three times higher than in those with controlled blood pressure. Similar findings were reported in Tigray,²⁹ Jimma,³² Spain²³ and the Republic of Korea.²⁴ This may be because uncontrolled blood pressure increases the risk of atherosclerosis, heart failure and stroke, thereby accelerating the progression of CKD, especially in those with existing renal impairment.

The odds of CKD in adult hypertensive patients with proteinuria at baseline were 4.48 times higher than in those without proteinuria. This finding is supported by the study in Gondar.³¹ This relationship may be due to uncontrolled HTN damaging the kidney’s glomeruli, leading to proteinuria. Excess protein in urine destroys glomeruli and tubules, causing inflammation that worsens kidney function and increases the risk of CKD.

Moreover, in hypertensive patients with high serum creatinine at baseline, the odds of CKD were 8.67 times higher than in those with normal serum creatinine. This is supported by the study in Spain²³ and Gondar.³¹ This may result from high blood pressure, which elevates serum creatinine levels and damages the blood vessels surrounding the kidneys. If underlying issues are not addressed, this condition can progress to chronic kidney impairment.

Furthermore, in adult hypertensive patients with dyslipidaemia at baseline, the odds of CKD were 1.72 times higher than in those without dyslipidaemia. The findings are supported by studies in Tigray²⁹ and Gondar.³¹ This association may result from dyslipidaemia exacerbating the buildup of fatty deposits in blood vessels, leading to renal artery stenosis. This condition compromises blood flow to the kidneys and contributes to CKD.⁶⁸

CKD is a significant public health issue that increases medical costs, mortality rates and hospital stays, while also presenting a poor prognosis. Developing a risk prediction model to identify high-risk groups is essential for implementing early preventive measures. This study developed a risk prediction model for CKD in patients over 18, demonstrating high predictive value and discrimination in internal validation. It is believed to be applicable in HTN management by identifying individuals at higher risk for CKD at UOGCSH.

The nomogram for predicting CKD included variables such as age, residence, family history of HTN, blood pressure status, type of HTN, baseline proteinuria, baseline serum creatinine level and dyslipidaemia. While individual predictors were effective, their combined effect was even stronger. A higher total score on the nomogram indicated a greater risk of CKD, and the model fit the data well. In the development dataset, the model demonstrated good prediction performance with an AUC of 91.99% (95% CI 90.07% to 93.91%), which was well-calibrated (p=0.367). Based on the prediction model performance classification, this prediction model had excellent discriminatory power.⁶⁹ The performance of this model was better than that of a study conducted in China, which developed a predictive model for hypertensive nephropathy with a concordance statistic equivalent to an AUC of 78.5%. Variables included were salt intake, diabetes mellitus, stroke, serum low-density lipoprotein, pulse pressure, age, HTN duration and serum uric acid.³³ Additionally, the performance of the nomogram in this study was better than the prediction risk of Chinese hypertensive patients, which had an AUC of 74% in the model development data.⁷⁰ Variables included in a study conducted in China were sex, age, smoking history, drinking history, coronary heart disease, diabetes history, C-reactive protein, Cystatin, β2-microglobulin protein, blood pressure type and renal artery resistance index. The variation in model performance might be due to the difference in the strength of the included predictor variables in the model development.

An optimal probability cut-off point of 0.1572 was identified, resulting in improved model performance, including better discrimination, sensitivity, specificity and predictive values. Using this probability as the cut-off point for CKD risk prediction, the model performance will have benefits for providing special attention and intervention for those identified as high-risk individuals. That means individuals with a probability of at least 0.1572 need close follow-up. The CKD risk prediction nomogram is designed to assist healthcare professionals in stratifying hypertensive patients based on their individual risk for CKD and providing targeted attention according to their risk levels. These findings inform users of the developed nomogram that early identification and providing close follow-up for high-risk hypertensive patients are cost-effective and impactful.⁷¹ However, the limitations of this study are as follows: first, its applicability in other settings is restricted due to a lack of external validation; therefore, the generalisability of the study is limited to different settings. Second, some important predictors like BMI, educational status, marital status, salt intake and socioeconomic status were not included because of the secondary nature of the data, which may limit the model’s accuracy.

As a result of this study, health professionals should use the developed model to identify high-risk patients for CKD by using these easily accessible prognostic predictors. Future researchers should conduct prospective studies to enhance the developed model by incorporating important predictors missed in the retrospectively collected data. Additionally, external validation should be performed before applying the model in another healthcare setting.

Policy implications and applications

The developed nomogram has policy implications and applications for advanced chronic illness in hypertensive patients, specifically for CKD, which has the worst outcome and requires the highest cost to get the treatment. First, the developed nomogram may be used by clinicians for early identification of high-risk hypertensive individuals for the incidence of CKD and to provide appropriate intervention during their follow-up. Second, the nomogram may be useful for monitoring changes in risk over time during follow-up. If risks are identified and treated early, they can reduce both personal financial costs and institutional expenses.

Conclusion

Our nomogram has shown excellent discrimination and calibration to identify hypertensive patients with a high risk of CKD. This model will help clinicians identify high-risk individuals early and provide better counselling and management with close follow-up.

Data availability statement

Data are available upon reasonable request.