Validation of the D:A:D chronic kidney disease risk score in people living with HIV: the IeDEA West Africa Cohort Collaboration

Abstract

Objectives

A risk score for long-term prediction of chronic kidney disease (CKD) in people living with HIV (PLHIV) has been developed using data from the D:A:D cohort. We assessed the performance of the D:A:D risk score in a cohort of PLHIV in West Africa.

Methods

Data from PLHIV starting antiretroviral treatment in four clinics in Burkina Faso, Côte d’Ivoire and Togo participating in the IeDEA West Africa collaboration were analysed. CKD was defined as two consecutive estimated glomerular filtration rates (eGFRs) of ≤ 60 mL/min/1.73 m². The D:A:D score (short version) was calculated using age, gender, nadir CD4 and baseline eGFR and was categorized into low, medium, and high-risk groups.

Results

In 14 930 participants (70% female, median age = 38 years; median nadir CD4 count = 183 cells/μL) followed for a median duration of 5.7 years, 660 (4.4%) progressed to CKD, with an incidence [95% confidence interval (CI)] of 7.8 (7.2–8.4) per 1000 person-years (PY). CKD incidence rates were 2.4 (2.0–2.8), 8.1 (6.8–9.6) and, 30.9 (28.0–34.1) per 1000 PY in the low-, medium- and high-risk groups, respectively. In the high-risk group, 14.7% (95% CI: 13.3; 16.3) had progressed to CKD at 5 years. Discrimination was good [C-statistics = 0.81 (0.79–0.83)]. In all, 79.4% of people who progressed to CKD were classified in the medium- to high-risk group at baseline (sensitivity) and 66.5% of people classified in the low risk group at baseline did not progress to CKD (specificity).

Conclusions

These findings confirm the validity of the D:A:D score in identifying individuals at risk of developing CKD who could benefit from enhanced kidney monitoring in West African HIV clinics.

Introduction

Before the advent of antiretroviral treatment (ART), kidney disease was a common complication of HIV infection. The introduction of ART has greatly reduced HIV-related kidney disease, such as HIV-related nephropathy (HIVAN), but chronic kidney disease (CKD) remains a major health concern in people living with HIV (PLHIV) in all parts of the world and particularly in sub-Saharan Africa, where people experience both high prevalence of HIV and high rates of CKD [1–4]. CKD is associated with increased cardiovascular morbidity and mortality and may progress to end-stage kidney disease, requiring kidney replacement therapy, the cost of which is high or even prohibitive in low-income countries. Thus, it is critical to identify people at risk of developing CKD to implement targeted intervention to prevent the deterioration of kidney function.

Mocroft et al. [5] have developed a simple prognostic risk score for CKD in HIV-infected individuals using data from the D:A:D cohort. An online tool for this score is available at https://www.chip.dk/Tools-Standards/Clinical-risk-scores. This score was externally validated in four cohorts from high-income countries, but has not yet been validated in cohorts in sub-Saharan Africa [5–7].

In this paper, we aimed to assess the performance of the D:A:D risk score to predict CKD in PLHIV in sub-Saharan Africa using data from the International Epidemiological Databases to Evaluate AIDS (IeDEA) West Africa Collaboration.

Methods

Source of data

The IeDEA West Africa Collaboration was created in 2006 as part of the global IeDEA collaboration of the US National Institutes of Health (https://www.iedea.org/regions/west-africa/). The database (seventh version) compiled data of HIV-infected individuals who had reached 16 years of age at ART initiation in seven HIV clinics from five countries (Benin, Burkina Faso, Ivory Coast, Togo and Senegal). Data were prospectively and retrospectively collected from 2006 to 2018.

In this analysis, we used data from four of the seven HIV outpatient clinics with comprehensive information on creatinine measurements. These sites included the University Teaching Hospital (CHU) Sourô Sanou in Bobo Dioulasso, Burkina Faso; the Centre de Prise en charge, de Recherche et de Formation (CePReF) and the Centre Intégré de Recherches Biocliniques d’Abidjan (CIRBA) in Abidjan, Côte d’Ivoire; and the University Teaching Hospital Sylvanus Olympio (CHUSO) in Lomé, Togo. Three clinics were not included because the validation of creatinine data was not completed.

Eligibility criteria

Our study population included individuals who initiated ART within 6 months of their first visit to the clinic and who had an estimated glomerular filtration rate (eGFR) > 60 mL/min/1.73 m² at baseline, given that baseline was defined as the first eGFR > 60 mL/min/1.73 m². Kidney function follow-up was based on creatinine measurements available during the individual’s follow-up in the clinic. Individuals with fewer than three eGFR measurements (including baseline) or with a follow-up of kidney function shorter than 3 months were excluded.

Outcome

In accordance with the D:A:D study, CKD was defined as two consecutive values of eGFR ≤ 60 mL/min/1.73 m², at least 3 months apart. eGFR was estimated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation without including the ethnic correction for black people, consistent with recent studies showing that this equation is more accurate without this ethnic correction in African populations [8,9].

Predictors

The short version of the D:A:D risk score is based on six predictors: age at baseline, gender, nadir CD4 count, eGFR baseline, hepatitis C and intravenous drug use. We assumed that no one had hepatitis C or was an intravenous drug user, as this information was not routinely recorded in the IeDEA cohort, mainly because these risk factors were not frequently encountered in this region. Missing values for nadir CD4 count were set as count ≤ 200 cells/μL (coefficient = 0). The score was categorized as low, medium or high as per cut-offs provided in the D:A:D cohort (< 0, 0–4, ≥ 5).

Statistical analysis

We assessed the predictive performance of the D:A:D risk score by examining measures of discrimination and calibration. Discrimination refers to the ability of the model to distinguish individuals who will develop CKD from those who will not. It was assessed with C-statistics for binary outcomes. Calibration, i.e. agreement between probabilities of outcome as predicted by the model and observed outcome frequencies, was assessed using the model-based approach proposed by Crowson et al. [10]. Predicted probability of progression to CKD during the follow-up of the individual was calculated using the exact coefficients of the model provided in Mocroft et al. (Table S1). Standardized incidence ratios were estimated using Poisson models. These ratios are equivalent to the ratio between observed and expected number of events according to the model, and were estimated for the total population and within risk score groups.

Sensitivity analyses were conducted to assess the validity of the model using the Modification of Diet in Renal Disease (MDRD) equation for non-standardized measures of creatinine to estimate kidney function and define CKD [11], and using CKD-EPI with the ethnic correction for black people. We also assessed the validity of the model in participants enrolled after 2010 when tenofovir disoproxyl fumarate (TDF) was first recommended as part of the preferred first-line antiretroviral regimen for HIV [12].

Finally, we evaluated the impact of TDF exposure at baseline on the risk of progression to CKD. Individuals were categorized as TDF-exposed if they were receiving TDF at baseline or if they started a TDF-based regimen within 6 months of the baseline visit. Incidence rate ratios for CKD by TDF exposure were estimated independently of the risk score category and also using a variable combining risk score category and TDF exposure.

We used Stata (v.14.0) for statistical analyses. We followed the recommendations for reporting a score validation provided in the ‘Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis’ (TRIPOD) statement [13].

Ethical statement

Data used in the IeDEA West Africa database have been prospectively collected as part of routine clinical care in the participating centre since 1998, with formal approvals from the local institutional review board and the US National Institutes of Health.

Results

Of 27 417 participants recorded in the database, 23 608 (86%) had at least one eGFR measurement > 60 mL/min/1.73 m² defined as baseline. Of these, 8678 (37%) individuals were excluded because they had fewer than three eGFR measurements or had < 3 months of follow-up, leaving 14 930 (54%) who met the eligibility criteria and were analysed further (Table S2).

The proportion of individuals included in the analysis varied between sites: 73% (n = 4903) of the individuals in the file of University Teaching Hospital Sourô Sanou in Bobo Dioulasso, Burkina Faso, were included; 59% (n = 5631) in CEPREF and 61% (3114) in CIRBA in Abidjan, Côte d’Ivoire; and only 21% (1282) in CHUSO in Lomé, Togo. In the latter clinic, follow-up was shorter with less frequent kidney assessment, and many participants did not verify eligibility criteria. Overall, included participants were more likely to be female, were enrolled more recently, were younger and had higher CD4 nadir and body mass index than non-included individuals. Median eGFR was also higher among included individuals (Table S2).

The baseline and follow-up characteristics of individuals included in the analysis are shown in Tables 1 and 2. The median age at baseline was 38.5 [interquartile range (IQR): 32.6–45.4] years; 35% were still ART-naïve at baseline and almost half of them had a nadir CD4 count < 200 cells/μL. The median (IQR) eGFR at baseline was 103 (87–116); 7.1% had a baseline eGFR of between 60 and 70 mL/min/1.73 m².

Table 1.

Baseline characteristics of included participants by clinic

	CHU Sourô Sanou		CePReF		CHUSO		CIRBA		Total		D:A:D cohort [5]
No. participants	4903		5631		1282		3114		14 930		17 954
Date of baseline visit, year	2010	(2007–2012)	2010	(2008–2012)	2011	(2010–2013)	2010	(2009–2013)	2010	(2009–2012)	2005	(2004–2007)
ART-naive
No	1088	22.2	1816	32.3	628	49.0	1668	53.6	5200	34.8	8117	45.2
Yes	3815	77.8	3815	67.7	654	51.0	1446	46.4	9730	65.2	9837	54.8
Sex
Male	1316	26.8	1413	25.1	412	32.1	1283	41.2	4424	29.6	13 130	73.1
Female	3587	73.2	4218	74.9	870	67.9	1831	58.8	10506	70.4	4824	26.9
Age (years)	37.2	(31.4–44.2)	38.0	(32.2–44.6)	39.3	(33.4–46.3)	41.2	(35.0–48.1)	38.5	(32.6–45.4)	40	(34–47)
≤ 35	1954	39.9	2080	36.9	397	31.0	777	25.0	5208	34.9
> 35 ≤ 50	2384	48.6	2838	50.4	676	52.7	1727	55.5	7625	51.1
> 50 ≤ 60	482	9.8	596	10.6	163	12.7	507	16.3	1748	11.7
> 60	83	1.7	117	2.1	46	3.6	103	3.3	349	2.3
eGFR (mL/min/1.73 m2)	98.6	(84.5–111.7)	110.2	(92.9–119.9)	87.9	(75.6–102.5)	105.0	(89.2–115.7)	103.4	(86.6–115.9)	104	(90–120)
> 60 ≤ 70	356	7.3	305	5.4	195	15.2	199	6.4	1055	7.1
> 70 ≤ 90	1309	26.7	915	16.2	488	38.1	612	19.7	3324	22.3
> 90	3238	66.0	4411	78.3	599	46.7	2303	74.0	10551	70.7
Nadir CD4 count (cells/μl)	176.0	(89.0–271.0)	184.0	(84.0–281.0)	167.0	(79.0–262.0)	224.6	(103.4–340.5)	183.0	(87.0–284.0)	290	(169–434)
≤ 200	1944	39.6	2521	44.8	420	32.8	838	26.9	5723	38.3
> 200	2659	54.2	3014	53.5	617	48.1	677	21.7	6967	46.7
missing	300	6.1	96	1.7	245	19.1	1599	51.3	2240	15.0
Body mass index (kg/m2)	20.9	(18.7–23.7)	20.3	(18.2–22.9)	22.2	(19.6–25.0)	22.6	(19.9–26.0)	20.8	(18.6–23.6)

CHU Sourô Sanou, University Teaching Hospital Sourô Sanou in Bobo Dioulasso, Burkina Faso; CePReF, the Centre de Prise en Charge, de Recherche et de Formation; CIRBA, Centre Intégré de Recherches Biocliniques d’Abidjan, Abidjan, Côte d’Ivoire; CHUSO, University Teaching Hospital Sylvanus Olympio, Lomé, Togo; ART, antiretroviral treatment; eGFR, estimated glomerular filtration rate.

Values are n (%) for categorical variables and median (interquartile range) for continuous variables.

Table 2.

Follow-up characteristics by clinic

	CHU Sourô Sanou		CePReF		CHUSO		CIRBA		Total		D:A:D cohort [5]
No. of participants	4903		5631		1282		3114		14 930		17 954
Follow-up duration (years)	6.3	(3.3–9.4)	5.5	(3.0–7.8)	4.7	(3.0–6.1)	5.8	(3.0–8.4)	5.7	(3.1–8.2)	6.1
Frequency of kidney assessments (per year)	2.0	(1.8–2.3)	1.8	(1.5–2.2)	0.8	(0.7–1.1)	2.1	(1.9–2.3)	1.9	(1.6–2.2)	3	(2–4)
Creatinine (μmol/L)	73.0	(64.0–83.0)	61.9	(53.0–79.6)	79.6	(70.7–88.4)	69.0	(57.5–83.1)	70.7	(60.6–81.0)
eGFR (mL/min/1.73 m2)	94.2	(81.1–106.5)	105.4	(87.8–115.2)	86.6	(74.0–101.1)	100.5	(85.0–111.7)	98.8	(83.4–111.2)
CKD
Developed CKD [n (%)]	233	4.8	227	4.0	43	3.4	157	5.0	660	4.4	641	3.6
Incidence of CKD [per 1000 PYFU (95% CI)]	7.7	(6.8–8.8)	7.2	(6.3–8.1)	7.5	(5.5–10.1)	9.3	(7.9–10.8)	7.8	(7.2–8.4)	6.2	(5.7–6.7)

CHU Sourô Sanou, University Teaching Hospital Sourô Sanou in Bobo Dioulasso, Burkina Faso; CePReF, the Centre de Prise en Charge, de Recherche et de Formation; CIRBA, Centre Intégré de Recherches Biocliniques d’Abidjan, Abidjan, Côte d’Ivoire; CHUSO, University Teaching Hospital Sylvanus Olympio, Lomé, Togo; ART, antiretroviral treatment; eGFR, estimated glomerular filtration rate; CI, confidence interval; CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; PYFU, person-years of follow-up.

Values are median (interquartile range) unless specified otherwise.

The median (IQR) kidney follow-up time (from baseline eGFR to last eGFR measurement) was 5.7 (3.1–8.2) years with some differences between clinics. The follow-up duration was shortest in CHUSO (4.7 years) and longest in CHU Sourô Sanou (6.3 years). There was a total of 153 888 eGFR measurements available with a median (IQR) frequency of 1.9 (1.6–2.2) measurements per year and per individual. A total of 660 (4.4%) participants developed a CKD during the follow-up, and the CKD incidence was 7.8 [95% confidence interval (CI): 7.2–8.4] per 1000 person-years of follow-up (PYFU). The highest incidence was observed in the CIRBA clinic with 9.3 per 1000 PYFU.

Compared with the D:A:D cohort, the IeDEA cohort enrolled a higher proportion of women. The IeDEA participants had lower nadir CD4 count, they were enrolled later but the median durations of follow-up were similar in both cohorts with more frequent kidney assessment in the D:A:D cohort (Tables 1 and 2).

The distribution of the risk score in the IeDEA cohort was similar to the distribution in the D:A:D cohort for the total population, but the median was lower in the IeDEA cohort than in the D:A:D cohort among people who developed CKD. Incidence increased markedly across the score categories (low, medium and high). At 5 years, 3.7% (95% CI: 3.4–4.0) had developed CKD, the proportion reaching 14.7% (13.3–16.3) for the high-risk group (Table 3).

Table 3.

Risk score model outcomes

	CHU Sourô Sanou		CePReF		CHUSO		CIRBA		Total		D:A:D cohort [5]
Baseline score [median (IQR)]	−2	(−5–4)	−2	(−5–1)	1	(−2–5)	−1	(−3–3)	−2	(−5–3)	−2	(−4–2)
Baseline score for those who developed CKD [median (IQR)]	6	(3–11)	5	(−1–10)	8	(5–11)	7	(1–11)	6	(1–11)	10	(5–13)
CKD incidence [per 1000 PYFU (95% CI)]
Low (score < 0)	2.24	(1.66–3.03)	2.50	(1.93–3.23)	1.57	(0.59–4.17)	2.75	(1.93–3.91)	2.42	(2.05–2.86)	0.56	(0.38–0.75)
Medium (score 0–4)	8.07	(6.07–10.75)	8.25	(6.03–11.30)	3.98	(1.79–8.87)	9.78	(6.84–13.99)	8.07	(6.76–9.64)	4.67	(3.80–5.53)
High (score ≥ 5)	28.09	(23.85–33.10)	35.75	(30.10–42.45)	19.48	(13.85–27.41)	37.14	(30.40–45.33)	30.91	(28.03–34.08)	36.05	(32.86–39.23)
Incidence rate ratio (95% CI)
Low (score < 0)	0.28	(0.18–0.42)	0.30	(0.20–0.45)	0.39	(0.11–1.39)	0.28	(0.17–0.46)	0.30	(0.23–0.38)	0.12	(0.08–0.18)
Medium (score 0–4)	1		1		1		1				1
High (score ≥ 5)	3.48	(2.50–4.84)	4.33	(3.03–6.20)	4.89	(2.05–11.67)	3.80	(2.52–5.72)	3.83	(3.13–4.69)	7.73	(6.29–9.49)
Kaplan-Meier [% progressed at 5 years (95% CI)]
Low (score < 0)	0.73	(0.46–1.17)	0.99	(0.69–1.41)	0.8	(0.3–2.14)	0.81	(0.47–1.40)	0.86	(0.67–1.10)	0.19	(0.10–0.27)
Medium (score 0–4)	3.90	(2.73–5.56)	4.10	(2.83–5.92)	1.76	(0.71–4.34)	3.51	(2.18–5.62)	3.63	(2.91–4.51)	1.62	(1.19–2.05)
High (score ≥ 5)	14.16	(11.86–16.85)	16.79	(13.97–20.10)	9.21	(6.45–13.08)	17.34	(13.95–21.44)	14.74	(13.27–16.35)	15.33	(13.82–16.84)
aIRR per unit increase score (95% CI)	1.23	(1.20–1.26)	1.24	(1.22–1.27)	1.21	(1.15–1.28)	1.22	(1.19–1.25)	1.23	(1.21–1.24)	1.32	(1.31–1.34)

CHU Sourô Sanou, University Teaching Hospital Sourô Sanou in Bobo Dioulasso, Burkina Faso; CePReF, the Centre de Prise en Charge, de Recherche et de Formation; CIRBA, Centre Intégré de Recherches Biocliniques d’Abidjan, Abidjan, Côte d’Ivoire; CHUSO, University Teaching Hospital Sylvanus Olympio, Lomé, Togo; aIRR, adjusted incidence rate ratio; CKD, chronic kidney disease; IQR, interquartile range; PYFU, person-years of follow-up; CI, confidence interval.

Incidence rates differed across the clinics, but were of the same order of magnitude. The adjusted incidence rate ratios (aIRRs) associated with an increase in the risk score of one point were similar in the four clinics and lower than in the D:A:D cohort.

Performance data of the D:A:D score are presented in Table 4. The C-statistics, ranging from 0.79 to 0.81, indicated an acceptable to good discrimination of the D:A:D score in all cohorts. Sensitivity and specificity at the cut-offs defining the three risk categories (≥ 0 and ≥ 5) were reported. When considering a cut-off at 0, 79.4% of people who progressed to CKD were correctly classified at baseline (i.e. score ≥ 0) (sensitivity) and 66.5% of people with a score < 0 at baseline did not progressed to CKD later (specificity). Sensitivity and specificity were 60.9% and 84.9%, respectively, for a cut-off at 5. The calibration parameters, corresponding to the ratio between observed number of events and expected number of events according to the model, indicated modest calibration of the model with under-estimation of the risk (parameter > 1) in all clinics but CHUSO, where calibration was 0.66 (95% CI: 0.49–0.89) (Table 4). Calibration by score categories indicated under-estimation for the low-risk group (which is the largest group) and over-estimation for medium- and high-risk categories (i.e. observed number of events were lower than expected).

Table 4.

Performance of the D:A:D risk score

	CHU Sourô Sanou		CePReF		CHUSO		CIRBA		Total
Discrimination
C-statistics	0.81	(0.78–0.84)	0.81	(0.78–0.84)	0.79	(0.73–0.86)	0.81	(0.77–0.85)	0.81	(0.79–0.83)
	≥ 0	≥ 5	≥ 0	≥ 5	≥ 0	≥ 5	≥ 0	≥ 5	≥ 0	≥ 5
Sensitivity	81.5	61.4	74.4	57.3	90.7	76.7	80.2	61.1	79.4	60.9
Specificity	64.4	83.6	73.5	89.1	44.3	71.3	66.3	85.1	66.5	84.9
Calibration
Calibration (O/E)
Low (score < 0)	2.28	(1.69–3.08)	2.52	(1.95–3.26)	1.50	(0.56–3.99)	2.45	(1.73–3.49)	2.38	(2.01–2.82)
Medium (score 0–4)	0.76	(0.50–1.15)	0.72	(0.48–1.08)	0.58	(0.16–2.06)	0.87	(0.53–1.43)	0.74	(0.58–0.94)
High (score ≥ 5)	0.43	(0.31–0.61)	0.49	(0.36–0.67)	0.40	(0.14–1.12)	0.45	(0.30–0.67)	0.43	(0.35–0.52)
Total	1.23	(1.08–1.39)	1.51	(1.33–1.72)	0.66	(0.49–0.89)	1.37	(1.18–1.61)	1.27	(1.18–1.37)

CHU Sourô Sanou, University Teaching Hospital Sourô Sanou in Bobo Dioulasso, Burkina Faso; CePReF, the Centre de Prise en Charge, de Recherche et de Formation; CIRBA, Centre Intégré de Recherches Biocliniques d’Abidjan, Abidjan, Côte d’Ivoire; CHUSO, University Teaching Hospital Sylvanus Olympio, Lomé, Togo; O/E, number of observed/expected events according to the D:A:D model.

Sensitivity analysis

The results obtained with the use of the MDRD equation to estimate eGFR led to poorer discrimination. Those obtained with the use of the CKD-EPI equation, including the ethnic correction, showed a lower incidence of CKD with somewhat higher discrimination performance of the score. The validation of the D:A:D score improved slightly when the enrolment date was restricted to individuals enrolled after 2010 when TDF was first recommended as part of the preferred first-line antiretroviral regimen for HIV and incidence rates of progression to CKD were higher (Table S3).

We evaluated the impact of receiving or starting a TDF-based regimen at baseline on the progression to CKD. Overall, 3798 (25.4%) individuals received or started TDF at baseline. TDF exposure was associated with a 60% increased risk of progressing to CKD (IRR = 1.6, 95% CI: 1.3–1.9). There was no interaction between score category and TDF exposure, indicating that the increases in risk were similar in each score category. An individual with a high-risk score and starting a TDF-based regimen had an incidence rate ratio of 19.8 (95% CI: 14.8–26.5) compared with individuals in the low-risk group not exposed to TDF (Fig. 1).

Fig. 1.

Initiation of a tenofovir disoproxyl fumarate (TDF)-based regimen and progression to chronic kidney disease (CKD). Individuals were classified as TDF-exposed (TDF) if they were receiving TDF at baseline or if they started a TDF-based regimen within 6 months of the baseline visit. Incidence rate ratios for progression to CKD were estimated using a variable combining risk score category and TDF exposure.

Discussion

In this analysis, we used a large cohort of about 15 000 PLHIV and followed in four HIV clinics in West Africa to assess the performance of the D:A:D short risk score for predicting CKD in the medium to long term. This score, based on basic information that is routinely available in HIV care in most resource-constrained settings, is readily calculable and can be used to identify people at risk to provide targeted monitoring and prevention intervention. To our knowledge, this is the first validation study of a score predicting CKD in HIV infection conducted in sub-Saharan Africa.

The D:A:D risk score showed acceptable to good discrimination in the IeDEA West Africa dataset with an overall C-statistic of 0.81 (95% CI: 0.79–0.83). We assessed sensitivity and specificity at the risk group cut-offs (0 and 5) and found that a threshold of 0 led to acceptable discrimination, with 79.4% of people who progressed to CKD during their follow-up classified in the at-risk category at baseline (sensitivity) and 66.4% of people classified in the low risk category (score < 0) not progressing to CKD (specificity). At a score of 5, specificity was very good but the sensitivity was poor and too many people at risk of developing CKD were not correctly classified.

The performance of the model with regard to discrimination was quite similar across the four HIV clinics. The CHUSO site displayed the lowest discrimination with a C-statistic of 0.79, but events were under-diagnosed at this clinic due to lower frequency of creatinine measurement and shorter follow-up. We also only assessed the performance of the model among participants enrolled in HIV care after 2010 because these individuals were better representatives of the current population. Indeed, in 2010, the eligibility criteria for ART as recommended by WHO guidelines were expanded from a CD4 count < 200 cells/μL to one < 350 cells/μL and TDF was first introduced as part of the preferred first-line option. Discrimination was good in this sub-population. As expected, TDF exposure at baseline was associated with increased risk of progression to CKD. However, in contrast with the findings from the D:A:D cohort, in our study the impact of TDF was not increased in the high-risk group.

Calibration parameters indicated overall under-estimation of the risk in three of the four sites. In CHUSO the predicted number of events was higher than the observed number and this supports the hypothesis that events were under-diagnosed at this site. Calibration across risk score groups showed substantial under-estimation in the low-risk group, while over-estimation was observed in the medium- and high-risk groups. This calibration issue was found consistently in all sites and may be partly explained by the lack of information on two predictors of the score, hepatitis C co-infection and intravenous drug use. Miscalibration of the model with overall under-estimation could also indicate that other important risk factors are missing and hypertension is one of them. In a recent study conducted in people initiating ART in the Bobo Dioulasso clinic, we found that hypertension was strongly associated with progression to CKD [14]. Another risk factor to consider is the APOL1-linked genetic predisposition. Two risk variants of the APOL1 gene, G1 and G2, were found to be strongly associated with HIVAN [15]. Interestingly, these risk variants are exclusively found among Africans and individuals of African descent, and several studies have reported a high prevalence of these mutations in West Africa [16]. This genetic factor may explain the higher incidence rate of CKD in the IeDEA compared with the US or European data.

Comparison with other external validation studies

To our knowledge, three scores have been developed for use in HIV infection to identify individuals with preserved kidney function (eGFR > 60 mL/min/1.73 m²) who are at increased risk of developing CKD later [5,17,18]. In addition to the D:A:D risk score, the Scherzer score was developed using data from HIV-positive male veterans in the US, including age, glucose, systolic blood pressure, hypertension, triglyceride level, proteinuria and CD4 cell count as predictors [18]. Another score was developed in a Japanese HIV cohort to predict CKD at 1 year using baseline eGFR, age, diabetes, proteinuria and CD4 count [17]. None of these scores has been validated in sub-Saharan African populations, although this part of the world hosts about two-thirds of PLHIV. We selected the D:A:D score for our validation study because it is based on limited data that can be made routinely available in most African settings. The D:A:D risk score was externally validated within the original publication using two datasets: the Royal Free Hospital Clinic Cohort [19], including individuals living in the UK, and the control arms of two international multicentric trials (SMART [20] and ESPRIT [21]) but with only 26 participants from sub-Saharan Africa. The short risk (in the Royal Free Hospital) as well as in the full risk score (in the trials) had good discrimination (C-statistic > 0.85). The D:A:D score was also evaluated using data from a large cohort in the US and from Australia [6,7], and the findings supported the validity of the score in predicting CKD in PLHIV in these settings. In the Australian cohort, the D:A: D score showed better discrimination performance than the Scherzer score [6]. Calibration was not formally evaluated in these reports.

Clinical implementation of the score

The D:A:D risk score could be easily implemented in HIV care in most African settings to identify the individuals who would benefit from closer CKD monitoring and prevention intervention. The use of the score could help to align resources with risk, with potential large benefits in settings where resources are constrained. We believe that the performance of the model is good enough to be used without recalibration at least in the initial stage. The performance of the model should first be assessed in datasets where all predictors are available before further updating if suboptimal calibration of the D:A:D model is confirmed in African settings. Addition of hypertension, a well-known risk factor of CKD, may significantly improve the performance of the model, as hypertension is highly prevalent and poorly controlled in Africa.

Strengths and limitations

There are several strengths to this study. First, its large sample size with data from four different sites located in three different countries in West Africa allowed us to highlight the stability and robustness of the results. Second, this study is based on data routinely collected in the medical files in standard HIV clinics, showing that the score is implementable in these settings. Several limitations should also be acknowledged. First, hepatitis C and intravenous drug use predictors were not consistently available and the score was calculated using only four out of six predictors for the short version of the risk score. In a recent systematic review, prevalence of HCV co-infection in people living with HCV was estimated at 3.6% in West Africa based on 11 studies. In Burkina Faso, the estimate reached 9.6% [22]. Second, we used the CKD-EPI equation to estimate eGFR; however, this equation should be used with standardized creatinine measurements [23]. Detailed information about laboratory methods used in the different sites was not available, and standardized creatinine measurements have only been available for a short time in these settings. To account for this, we assessed the performance of the model using the MDRD equation for non-standardized creatinine measurements and found that the model performed slightly less well than with CKD-EPI equation. Although the CKD-EPI equation is generally preferred because it estimates GFR more accurately [24], several reports have shown that the performance of a score for CKD prediction or diagnosis was not influenced by the choice of the equation [5,25]. Finally, about half of the population did not verify the eligibility criteria and were therefore excluded from the analysis. As those who were excluded had lower nadir CD4 and lower eGFR, we cannot exclude the contribution of a selection or survival bias for people at high risk. Yet the fact that the risk score performed well, even in a population at lower risk (i.e. one that included participants with higher eGFR and CD4 count), strengthens our results.

Conclusion

The performance of the D:A:D score in predicting CKD using four predictors was acceptable. This score, based on limited information that is routinely available in HIV care in most resource-constrained settings, is readily calculable and can be used to identify people at highest risk to provide targeted monitoring and prevention intervention. We found that the threshold of zero can be used to identify those individuals who are most likely to benefit from closer monitoring of kidney function in order to prevent progression to CKD. This will help to align risk and resources in resource-constrained settings of sub-Saharan Africa.

Supplementary Material

STable_1_3

Table S1. Models for short risk score for CKD provided by Mocroft et al. [5].

Table S2. Characteristics of participants excluded and included in the analysis.

Table S3. Sensitivity analyses for risk score model outcomes (MDRD equation for GFR estimation and post-2010 period).