Risks of Major Adverse Kidney Events in Non-Hispanic Black Patients with Diabetes or Hypertension: A Real-World Cohort Study

Abstract

Introduction

Chronic kidney disease (CKD) disproportionately burdens non-Hispanic Black (NHB) patients who experience a three- to four-fold higher risk of kidney failure than non-Hispanic White (NHW) individuals. Diabetes and hypertension are the leading causes of CKD, yet the influence of race, social context, and clinical factors on major adverse kidney events (MAKEs) remains inadequately understood. This study examined relationships between race, social factors, and MAKE among patients with diabetes or hypertension.

Methods

Electronic health record data from the CURE-CKD Registry (2013–2022) were used to assemble two mutually exclusive cohorts of NHB and NHW (reference) adults with (1) diabetes (N = 375,605) or (2) hypertension without diabetes (N = 710,768). The primary outcome was time to first MAKE, defined as ≥40% decline in estimated glomerular filtration rate (eGFR), eGFR <15 mL/min/1.73 m², kidney failure, dialysis, transplant, or death. Cox proportional hazards models estimated associations between predictors and MAKE. Extreme gradient boosting (XGBoost) machine learning (ML) models were used as a complement to the traditional survival models to evaluate variable relationships with MAKE and compare predictive performance.

Results

NHB patients were younger (diabetes: 56 ± 15 vs. 62 ± 14 years; hypertension: 50 ± 16 vs. 59 ± 16 years), had more prevalent CKD (diabetes: 19% vs. 15%; hypertension: 8% vs. 7%), and had higher adjusted MAKE risk than NHW patients (diabetes: HR = 1.07, 95% CI: 1.04–1.11; hypertension: HR = 1.10, 95% CI: 1.06–1.14). MAKE occurred in 24% (n = 90,910) of the diabetes population over a median (interquartile range) of 4.1 (2.0–6.3) years, and 15% (n = 106,209) of the hypertension population over 4.4 (2.4–6.3) years. Non-commercial insurance (Medicaid, Medicare, or unclassified insurance [reference: Commercial]), higher social vulnerability index, more frequent hospitalizations, and urban residence were the top predictors of MAKE. XGBoost outperformed Cox models for MAKE prediction and SHapley Additive exPlanations value rankings confirmed key predictors identified in Cox models.

Conclusion

Across two large cohorts for CKD, NHB patients experienced higher MAKE risk, shaped by neighborhood social and structural conditions rather than demographic and clinical characteristics alone. Integrating social risk assessment, improving guideline-directed testing, and using complementary ML and Cox modeling approaches may enhance early identification of CKD risk, support targeted intervention, and reduce disparities in CKD outcomes.

Introduction

Chronic kidney disease (CKD) has become a public health crisis in the United States (US) and worldwide. In the year 2023, CKD affected 40 million Americans with at least 147,000 attributable deaths [1–3]. If current trends continue, CKD will become the fifth leading cause of death globally by 2040 [4, 5]. Diabetes and hypertension are the primary risk factors for CKD and account for the majority of CKD cases in the US, disproportionately impacting racial and ethnic minority (minoritized) populations [6]. Among racially minoritized populations in the US, non-Hispanic Black (NHB) patients experience a three- to four-fold increased risk of developing end-stage kidney disease compared to non-Hispanic White (NHW) patients [1, 7–11]. These differences reflect a complex interplay of social, economic, systemic and biologic factors that may exacerbate the risks of diabetes and hypertension for major adverse kidney events (MAKEs) [1, 12–16]. For prognostic purposes, MAKE is comprised of estimated glomerular filtration rate (eGFR) decline of ≥40% or more, eGFR <15 mL/min/1.73 m² or initiation of kidney replacement therapy, and death [17]. Important gaps remain in understanding MAKE incidence or risk prediction among at-risk groups, such as those with diabetes or hypertension, particularly in relation to race and social factors.

Given that diabetes and hypertension account for the largest populations with, and at risk for CKD and kidney failure, focusing on these groups enables targeted risk stratification and advances understanding of how clinical and social factors contribute to differences in CKD outcomes. Therefore, the study aim was to assess these factors for associations with MAKE in patients with diabetes or hypertension by NHB or NHW race. Additionally, both traditional statistical models and machine learning (ML) methods evaluated risk factors and compared predictive performance between modeling approaches.

Methods

Data Source

Two study cohorts (diabetes and non-diabetic hypertension) were sourced from the Center for Kidney Disease Research, Education, and Hope (CURE-CKD) Registry [18]. Demographic, clinical, and social factors were extracted from electronic health records (EHR) from Providence, a large health system in the Western US. This study was conducted according to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines for cohort studies [19].

Study Populations

Adult (≥18 years) NHB and NHW populations were identified as having either diabetes or hypertension without concurrent diabetes (hereafter referred to as “hypertension”), during 2013–2022, forming two cohorts. Diabetes was defined using laboratory criteria (hemoglobin-HbA1c ≥6.5% or two elevated blood glucose values), ICD-10 diagnosis codes (one inpatient or two outpatient), or prescriptions for a glucose-lowering medication excluding inpatient insulin, off-label metformin use associated with a polycystic ovarian syndrome diagnosis, or other medications used to treat non-diabetes conditions such as glucagon-like peptide-1 (GLP-1) receptor agonists and sodium-glucose cotransporter-2 (SGLT2) inhibitors. Hypertension was defined using repeated elevated blood pressure measurements at outpatient encounters (two systolic ≥140 mm Hg or diastolic ≥90 mm Hg measures ≥14 days apart) or an ICD-10 diagnosis code. For the diabetes cohort, exclusion criteria included prior kidney failure, a primary CKD diagnosis (not attributed to diabetes or hypertension, online suppl. Table S1; for all online suppl. material, see https://doi.org/10.1159/ajn/ablag006), missing eGFR measures in baseline or follow-up periods, and missing social factor data (social vulnerability index [SVI], rural-urban commuting area). For the hypertension cohort, additional exclusion criteria included missing baseline blood pressure measurements and evidence of diabetes before follow-up. The SVI, developed by the Centers for Disease Control and Prevention, is a composite measure derived from 15 American Community Survey-based indicators across four domains: socioeconomic status (e.g., income, poverty, employment), household composition and disability (e.g., age, disability status), minority status and language, and housing type and transportation [20]. The SVI provides a 0–1 census-tract score with higher values indicating greater social vulnerability [20]. Insurance type was collected from encounters at entry and categorized as Medicare (a US federal insurance program providing coverage for older adults and individuals with disabilities), Medicaid (a US federal insurance program providing coverage for low-income populations), Commercial (employer-sponsored or private insurance), or unclassified. Analyses were conducted using complete cases for both cohorts, and individuals with missing data for key covariates were excluded (online suppl. Fig. S1). Serum creatinine measurements were used to calculate eGFR using the 2021 CKD epidemiology collaboration equation [21].

Study Design and Outcomes

For this real-world longitudinal cohort study, the index date was defined as 90 days after the identification of diabetes or hypertension to optimize baseline data availability by allowing time for data to accrue in the EHR after entry into the study. Baseline data were accrued 1 year prior to the index date. The primary outcome of this study was time to first MAKE, defined as ≥40% eGFR decline from baseline, eGFR <15 mL/min/1.73 m², administrative codes for kidney failure, dialysis, or transplant, or all-cause death ascertained from both the Social Security Death Index and the EHR. Patients were followed through their first MAKE or last encounter through December 31, 2022, whichever occurred first (online suppl. Fig. S2).

Statistical Analysis

Characteristics of patients with diabetes and hypertension were stratified by racial and ethnic identity. Frequency distributions were reported for categorical variables, median and interquartile range (IQR) were reported for skewed continuous variables, and the mean and standard deviation (SD) were reported for normally distributed continuous variables. The Kaplan-Meier survival estimates for MAKE were stratified by racial group and compared by log-rank test.

Cox proportional hazards models were paired with extreme gradient boosting (XGBoost) ML models to evaluate variable relationships with MAKE and compare performance for prediction of MAKE by average precision, area under the receiver operating characteristic curve (AUC-ROC), precision and Brier Score. The XGBoost model was selected after comparison to other binary classifiers; logistic regression and random forest, which demonstrated the best predictive performance based on the highest average precision score (online suppl. Methods; online suppl. Tables S2–S4; online suppl. Fig. S3).

Fully adjusted models included the same set of pre-selected predictors for each cohort based on clinical insights: demographics (age [modeled continuously], sex, race [reference: NHW]), social factors (urban residence [reference: rural], SVI [score scaled from 0 to 1], insurance type [categorized as Medicare, Medicaid, or unclassified, reference: Commercial]), health system encounters (number of nephrology care visits and number of hospitalizations), clinical variables (mean eGFR per −10 mL/min/1.73 m² and hypertension status at baseline [diabetes cohort only]), and medication use (angiotensin-converting enzyme [ACE] inhibitors or angiotensin receptor blockers [ARBs]). Models for the hypertension cohort additionally included prescriptions for nonsteroidal anti-inflammatory drugs (NSAIDs) and proton pump inhibitors (PPIs) because of the possible association between these medication classes and hypertension.

Cox proportional hazards models were fit to estimate both unadjusted and adjusted hazard ratios (HRs) for the prediction of MAKE. The assumption of proportional hazards was evaluated with Schoenfeld residual plots. p values <0.05 were considered statistically significant. Cox analyses were performed in R version 4.4.2 with the Survival package version 3.7.0. ML analyses were performed in Python version 3.10.8 with the Model Tuner package version 0.0.27b (online suppl. Methods). SHapley Additive exPlanations (SHAP) values were generated with the Tree SHAP algorithm implemented in the SHAP library (version 0.46.0) based on the trained XGBoost model [22–26]. The importance of features for the prediction of MAKE was ranked using the average absolute SHAP values, while individual SHAP values were used to visualize the direction of effect on MAKE prediction.

Subgroup and Exploratory Analyses

Cox proportional hazards models from the main analysis were fit for the NHB population in each cohort, diabetes or hypertension, to assess for potential differences in relative effect size or statistical significance compared to the total study population. As an exploratory ML analysis, urine albumin-to-creatinine ratio (UACR) and urine protein-to-creatinine ratio (UPCR) were included in the models after predictive mean matching imputation. We evaluated these exploratory models using the same predictive performance metrics as described in the Supplemental Methods.

Results

Baseline Characteristics – Diabetes Cohort

The diabetes cohort included 375,605 patients, of which 6% (n = 21,322) were NHB and 94% (n = 354,283) were NHW. The NHB population was younger (mean ± SD: 56 ± 15 years) than the NHW population (62 ± 14 years). More patients lived in urban than rural areas, with urban residence being more common in the NHB group (NHB: 95%, n = 20,195; NHW: 77%, n = 273,266). Most NHB patients had commercial (33%, n = 7,115) or unclassified insurance (28%, n = 5,864) and most NHW patients had commercial (31%, n = 110,929) or Medicare insurance (39%, n = 136,914). The mean ± SD SVI was 0.65 ± 0.27 among NHB patients and 0.50 ± 0.26 among NHW patients. At baseline, both race groups had similar mean ± SD eGFR, 82 ± 24 mL/min/1.73 m² in the NHB group and 83 ± 22 mL/min/1.73 m² in the NHW group. The proportion of patients with hypertension was 32% (n = 6,758) among NHB patients and 28% (n = 99,539) among NHW patients. Regardless of race, only 41% (n = 154,855) of patients were prescribed an ACE inhibitor or ARB medication (Table 1).

Table 1.

Baseline characteristics of the diabetes population in 2013–2022

​	Overall	NHB	NHW (reference)
Demographics
Persons, n (% of overall)	375,605 (100)	21,322 (6)	354,283 (94)
Sex, n (%)
Male	183,335 (49)	9,768 (46)	173,567 (49)
Female	192,270 (51)	11,554 (54)	180,716 (51)
Age, mean±SD, years	62±14	56±15	62±14
Age category, n (%)
18–39	33,926 (9)	3,355 (16)	30,571 (9)
40–59	113,620 (30)	9,053 (42)	104,567 (30)
60–79	184,158 (49)	7,678 (36)	176,480 (50)
≥80	43,901 (12)	1,236 (6)	42,665 (12)
Residence measures
Social Vulnerability Index, mean±SD	0.51±0.26	0.65±0.27	0.50±0.26
Socioeconomic status index	0.45±0.26	0.59±0.27	0.44±0.25
Housing composition index	0.50±0.27	0.57±0.29	0.50±0.27
Minority status index	0.46±0.22	0.70±0.21	0.45±0.21
Housing type	0.58±0.28	0.66±0.27	0.57±0.28
Dwelling area, n (%)
Rural	82,144 (22)	1,127 (5)	81,017 (23)
Urban	293,461 (78)	20,195 (95)	273,266 (77)
Health care and care patterns
Primary health insurance, n (%)
Commercial	118,044 (31)	7,115 (33)	110,929 (31)
Medicaid	29,168 (8)	3,384 (16)	25,784 (7)
Medicare	141,873 (38)	4,959 (23)	136,914 (39)
Unknown	86,520 (23)	5,864 (28)	80,656 (23)
Hospitalization, n (%)	135,824 (36)	8,018 (38)	127,806 (36)
Primary care encounter, n (%)	353,932 (94)	18,952 (89)	334,980 (95)
Internal medicine encounter, n (%)	78,602 (21)	3,224 (15)	75,378 (21)
Family medicine encounter, n (%)	153,099 (41)	9,425 (44)	143,674 (41)
Nephrology encounter, n (%)	4,581 (1)	262 (1)	4,319 (1)
Endocrinology encounter, n (%)	15,749 (4)	914 (4)	14,835 (4)
Clinical
CKD, n (%)	55,617 (15)	4,025 (19)	51,592 (15)
eGFR, n (%)	375,605 (100)	21,322 (100)	354,283 (100)
Mean±SD, mL/min/1.73 m2	82.5±22.2	81.5±24.1	82.6±22.1
eGFR category, n (%)
≥60	315,706 (84)	17,391 (82)	298,315 (84)
45–59	37,580 (10)	2,286 (11)	35,294 (10)
30–44	17,067 (5)	1,153 (5)	15,914 (4)
15–29	5,252 (1)	492 (2)	4,760 (1)
UACR, n (%)	43,443 (6)	3,436 (16)	40,007 (11)
Median (IQR), mg/g	12.0 (6.0–36.0)	12.0 (5.0–40.3)	12.0 (6.0–35.0)
UPCR, n (%)	2,262 (1)	249 (1)	2,013 (1)
Median (IQR), g/g	0.20 (0.10–0.47)	0.17 (0.10–0.44)	0.20 (0.10–0.47)
UACR or UPCR category, n (% of measured)
<30 mg/g or <0.15 g/g	31,858 (70)	2,472 (68)	29,386 (71)
30–300 mg/g or 0.15–0.5 g/g	10,610 (23)	883 (24)	9,727 (23)
>300 mg/g or >0.5 g/g	2,821 (6)	283 (8)	2,538 (6)
Hypertension, n (%)	106,297 (28)	6,758 (32)	99,539 (28)
Systolic blood pressure, outpatient, n (%)	310,922 (83)	16,906 (79)	294,016 (83)
Mean±SD, mm Hg	129±17	134±18	129±17
HbA1c, n (%)	220,202 (59)	13,465 (63)	206,737 (58)
Mean±SD, %	7.0±1.8	7.5±2.2	7.0±1.8
BMI, n (%)	312,016 (83)	17,004 (80)	295,012 (83)
Mean±SD, kg/m2	32±8	32±8	32±8
Medications, n (%)
ACE inhibitor or ARB	154,855 (41)	8,738 (41)	146,117 (41)
GLP-1 receptor agonist	6,647 (2)	393 (2)	6,254 (2)
SGLT2 inhibitor	5,404 (1)	323 (2)	5,081 (1)

SD, standard deviation; CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; UACR, urine albumin-to-creatinine ratio; IQR, interquartile range; UPCR, urine protein-to-creatinine ratio; Hb, hemoglobin; BMI, body mass index; ACE, angiotensin converting enzyme; ARB, angiotensin receptor blocker; GLP-1, glucagon-like peptide; SGLT2, sodium-glucose co-transporter.

Predictors of MAKE – Diabetes Cohort

A total of 24% (n = 90,910) of patients in the diabetes cohort experienced a MAKE outcome over a median (IQR) follow-up of 4.1 (2.0–6.3) years. A ≥40% decline in eGFR was observed in 84% (n = 4,167) of NHB and 72% (n = 62,043) of NHW patients who experienced MAKE. NHB patients had lower MAKE survival over 10 years compared to NHW patients (p = 0.006; Table 2; Fig. 1a).

Table 2.

Summary of unadjusted MAKE among the diabetes population

​	Overall	NHB	NHW (reference)
	N = 375,605	N = 21,322	N = 354,283
MAKE, n (%)	90,910 (24)	4,990 (23)	85,920 (24)
Event type, n (% of MAKE)
eGFR ≥40% decline	66,210 (73)	4,167 (84)	62,043 (72)
Kidney failure	15,235 (17)	1,236 (25)	13,999 (16)
Dialysis	6,614 (7)	592 (12)	6,022 (7)
Transplant	296 (0.3)	29 (0.6)	267 (0.3)
All-cause death	42,409 (47)	1,597 (32)	40,812 (48)

MAKE, major adverse kidney event; eGFR, estimated glomerular filtration rate.

Fig. 1.

Kaplan-Meier survival analysis for time to MAKE by race. a Diabetes cohort (N = 375,605). b Hypertension cohort (N = 710,768). The red line represents the NHW population; the blue line represents the NHB population. MAKE, major adverse kidney event; NHW, non-Hispanic White; NHB, non-Hispanic Black.

In the adjusted Cox model, insurance type was the strongest predictor of MAKE. Compared to commercial insurance, the hazard was highest among patients with Medicaid (HR = 1.98; 95% confidence interval [CI]: 1.92–2.04, p < 0.001), followed by unclassified insurance (HR = 1.85; 95% CI: 1.81–1.89, p < 0.001) and Medicare (HR = 1.42; 95% CI: 1.39–1.45, p < 0.001). Each 0.1-unit increase in the SVI was associated with a 4% higher hazard of MAKE (HR = 1.04; 95% CI: 1.04–1.05; p < 0.001), while urban residence was not significantly associated with MAKE hazard. Men had a higher hazard than women (HR = 1.08; 95% CI: 1.07–1.10, p < 0.001), and hazard increased modestly with age (HR = 1.02 per year; 95% CI: 1.02–1.03, p < 0.001). NHB patients had a slightly higher adjusted hazard of experiencing a MAKE outcome compared to NHW patients (HR = 1.07; 95% CI: 1.04–1.11; p < 0.001). Evidence of hypertension (HR = 1.08, 95% CI: 1.06–1.09, p < 0.001), prescription for ACE inhibitor or ARB medications (HR = 1.14, 95% CI: 1.12–1.15, p < 0.001), lower baseline eGFR (HR = 1.13, 95% CI: 1.12–1.13, p < 0.001), more hospitalizations (HR = 1.31, 95% CI: 1.31–1.32, p < 0.001), and a higher number of nephrology visits (HR = 1.02, 95% CI: 1.02–1.03, p < 0.001) were associated with greater hazard of MAKE (Fig. 2a; online suppl. Table S5).

Fig. 2.

Predictors of MAKE. a Diabetes cohort (N = 375,605). b Hypertension cohort (N, 710,768). MAKE, major adverse kidney event; HR, hazard ratio; CI, confidence interval; NHB, non-Hispanic Black; NHW, non-Hispanic White; eGFR, estimated glomerular filtration rate; ACE, angiotensin converting enzyme; ARB, angiotensin receptor blocker; PPI, proton pump inhibitor; NSAID, nonsteroidal anti-inflammatory drug.

The XGBoost model demonstrated better predictive performance than the Cox models, with significantly higher average precision (XGBoost 0.457; Cox 0.439, p < 0.001) and AUC-ROC (XGBoost 0.727; Cox 0.710, p < 0.001, online suppl. Table S2). In the SHAP analysis of the XGBoost model, higher hospitalization counts, lack of commercial insurance, older age, lower baseline eGFR, and higher SVI, were the top five most influential predictors of the MAKE outcome. These findings were consistent with the Cox proportional hazard models, which also identified hospitalization burden, insurance status, baseline eGFR, and social vulnerability as key predictors. However, differences emerged in the relative importance of predictors and how they contributed to the model. Further, Medicaid and Medicare insurance types, counts of nephrology care visits, urban residence, male sex, and race were features of little to no importance to the prediction of MAKE, though they were significant in the Cox model (online suppl. Fig. S4, S5).

In the subgroup analysis of the NHB-only population, Cox proportional hazards model insights were largely consistent with the overall diabetes cohort. Higher hazards were observed for Medicare (HR = 1.65; 95% CI: 1.51–1.81, p < 0.001) and unclassified insurance (HR = 2.05; 95% CI: 1.88–2.23, p < 0.001) compared to commercial insurance. Urban residence was significantly protective in the NHB subgroup (HR = 0.77; 95% CI: 0.69–0.86, p < 0.001), unlike in the full cohort. Hypertension (HR = 1.15, 95% CI: 1.08–1.23, p < 0.001) and ACE inhibitor or ARB use (HR = 1.19, 95% CI: 1.12–1.27, p < 0.001) remained significant predictors of increased hazard while nephrology visits were not significant (Fig. 3a; online suppl. Table S6).

Fig. 3.

Predictors of MAKE in the non-Hispanic Black-only. a Diabetes cohort (N = 21,322). b Hypertension cohort (N = 32,739). MAKE, major adverse kidney event; HR, hazard ratio; CI, confidence interval; eGFR, estimated glomerular filtration rate; ACE, angiotensin converting enzyme; ARB, angiotensin receptor blocker; PPI, proton pump inhibitor; NSAID, nonsteroidal anti-inflammatory drug.

Baseline Characteristics – Hypertension Cohort

The hypertension cohort included 710,768 patients; 5% (n = 32,739) were NHB and 95% (n = 678,029) were NHW. Like the diabetes cohort, NHB patients with hypertension were younger (mean ± SD: 50 ± 16 years) than NHW (59 ± 16 years) patients. The cohort consisted of more women (56%, n = 394,598) than men (44%, n = 316,170). Most patients, regardless of racial identity, lived in urban areas (80%, n = 566,077) compared to rural areas (20%, n = 144,691). In the NHB group, most patients had commercial (40%, n = 13,244) or Medicaid insurance (23%, n = 7,487), while most patients in the NHW group had either commercial (37%, n = 248,522) or Medicare (32%, n = 214,322) insurance. The mean ± SD SVI score was 0.63 ± 0.27 among NHB patients and 0.47 ± 0.26 among NHW patients. At baseline, the mean ± SD eGFR was 88 ± 22 mL/min/1.73 m² in the NHB group and 87 ± 20 mL/min/1.73 m² in the NHW group. Few patients regardless of racial identity group were prescribed ACE inhibitor or ARB medications (25%, n = 174,362; Table 3).

Table 3.

Baseline characteristics of the hypertension population in 2013–2022

​	Overall	NHB	NHW (reference)
Demographics
Persons, n (% of overall)	710,768 (100)	32,739 (5)	678,029 (95)
Sex, n (%)
Male	316,170 (44)	14,040 (43)	302,130 (45)
Female	394,598 (56)	18,699 (57)	375,899 (55)
Age, mean±SD, years	59±16	50±16	59±16
Age category, n (%)
18–39 years	114,643 (16)	9,909 (30)	104,734 (15)
40–59 years	225,025 (32)	13,762 (42)	211,263 (31)
60–79 years	292,689 (41)	7,840 (24)	284,849 (42)
≥80 years	78,411 (11)	1,228 (4)	77,183 (11)
Residence measures
Social Vulnerability Index, mean±SD	0.48±0.26	0.63±0.27	0.47±0.26
Socioeconomic status index	0.43±0.25	0.58±0.27	0.42±0.25
Housing composition index	0.48±0.27	0.54±0.29	0.47±0.27
Minority status index	0.46±0.21	0.68±0.21	0.45±0.21
Housing type	0.56±0.29	0.65±0.27	0.56±0.29
Dwelling area, n (%)
Rural	144,691 (20)	1,479 (5)	143,212 (21)
Urban	566,077 (80)	31,260 (95)	534,817 (79)
Health care and care patterns
Primary health insurance, n (%)
Commercial	261,766 (37)	13,244 (40)	248,522 (37)
Medicaid	70,591 (10)	7,487 (23)	63,104 (9)
Medicare	219,513 (31)	5,191 (16)	214,322 (32)
Unknown	158,898 (22)	6,817 (21)	152,081 (22)
Hospitalization, n (%)	160,521 (23)	6,821 (21)	153,700 (23)
Primary care encounter, n (%)	666,115 (94)	28,312 (87)	637,803 (94)
Internal medicine encounter, n (%)	148,632 (21)	4,494 (14)	144,138 (21)
Family medicine encounter, n (%)	316,485 (45)	15,149 (46)	301,336 (44)
Nephrology encounter, n (%)	4,308 (1)	184 (1)	4,124 (1)
Endocrinology encounter, n (%)	14,382 (2)	511 (2)	13,871 (2)
Clinical
CKD, n (%)	52,864 (7)	2,695 (8)	50,169 (7)
eGFR, n (%)	710,768 (100)	32,739 (100)	678,029 (100)
Mean±SD, mL/min/1.73 m2	86.9±20.4	88.2±21.8	86.9±20.3
eGFR category, n (%)
≥60	641,754 (90)	29,584 (90)	612,170 (90)
45–59	49,461 (7)	2,265 (7)	47,196 (7)
30–44	16,207 (2)	682 (2)	15,525 (2)
15–29	3,346 (0.5)	208 (0.6)	3,138 (0.5)
UACR, n (%)	18,099 (3)	1,546 (5)	16,553 (2)
Median (IQR), mg/g	8.0 (4.0–21.0)	7.0 (4.0–19.0)	8.0 (4.0–21.0)
UPCR, n (%)	3,616 (0.5)	354 (1)	3,262 (0.5)
Median (IQR), g/g	0.14 (0.10–0.25)	0.13 (0.09–0.26)	0.14 (0.10–0.25)
UACR or UPCR category, n (% of measured)
<30 mg/g or <0.15 g/g	16,328 (76)	1,437 (77)	14,891 (76)
30–300 mg/g or 0.15–0.5 g/g	4,123 (19)	338 (18)	3,785 (19)
>300 mg/g or >0.5 g/g	897 (4)	94 (5)	803 (4)
Systolic blood pressure, outpatient, n (%)	710,768 (100)	32,739 (100)	678,029 (100)
Mean±SD, mm Hg	134±17	137±17	134±16
HbA1c, n (%)	116,196 (16)	7,216 (22)	108,980 (16)
Mean±SD, %	5.6±0.4	5.6±0.4	5.6±0.4
BMI, n (%)	664,394 (93)	29,648 (91)	634,746 (94)
Mean±SD, kg/m2	29±7	31±8	29±7
Medications, n (%)
Medication use, n (%)
ACE inhibitor or ARB	174,362 (25)	7,130 (22)	167,232 (25)
GLP-1 receptor agonist	362 (0.1)	20 (0.1)	342 (0.1)
SGLT2 inhibitor	302 (0.04)	14 (0.04)	288 (0.04)
NSAID	264,361 (37)	14,316 (44)	250,045 (37)
PPI	167,065 (24)	6,406 (20)	160,659 (24)

SD, standard deviation; CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; UACR, urine albumin-to-creatinine ratio; IQR, interquartile range; UPCR, urine protein-to-creatinine ratio; Hb, hemoglobin; BMI, body mass index; ACE, angiotensin converting enzyme; ARB, angiotensin receptor blocker; GLP-1, glucagon-like peptide; SGLT2, sodium-glucose co-transporter; NSAID, nonsteroidal anti-inflammatory drug; PPI, proton pump inhibitor.

Predictors of MAKE – Hypertension Cohort

In the hypertension cohort, 15% (n = 106,209) experienced a MAKE outcome over a median (IQR) follow-up of 4.4 (2.4–6.3) years. Among those, 80% (n = 3,063) of NHB and 66% (n = 67,058) of NHW patients had a ≥40% decline in eGFR. Unadjusted survival differences by race group were also significant (p < 0.001) with NHW patients having lower overall MAKE survival than NHB patients (Table 4; Fig. 1b).

Table 4.

Summary of unadjusted MAKE among the hypertension population

​	Overall	NHB	NHW (reference)
	N = 710,768	N = 32,739	N = 678,029
MAKE, n (%)	106,209 (15)	3,850 (12)	102,359 (15)
Event type, n (%)
eGFR ≥40% decline	70,121 (66)	3,063 (80)	67,058 (66)
Kidney failure	11,374 (11)	649 (17)	10,725 (10)
Dialysis	4,542 (4)	283 (7)	4,259 (4)
Transplant	244 (0.2)	16 (0.4)	228 (0.2)
All-cause death	53,903 (51)	1,261 (33)	52,642 (51)

MAKE, major adverse kidney event; eGFR, estimated glomerular filtration rate.

In the unadjusted Cox model, NHB patients were less likely to experience MAKE (HR = 0.83; 95% CI: 0.81–0.86; p < 0.001) (online suppl. Table S7), but after adjustment for social factors and clinical covariates, NHB patients were more likely to experience MAKE than NHW patients (HR = 1.10; 95% CI: 1.06–1.14; p < 0.001). In the adjusted Cox model, insurance type was the strongest predictor. Compared to commercial insurance, MAKE risk was significantly higher for patients with Medicaid (HR = 2.51, 95% CI: 2.44–2.58, p < 0.001), Medicare (HR = 1.46, 95% CI: 1.43–1.49, p < 0.001), and unclassified insurance (HR = 1.50, 95% CI: 1.46–1.53, p < 0.001). Each 0.1-unit increase in SVI was associated with a 5% increased hazard of MAKE (HR = 1.05; 95% CI: 1.05–1.05; p < 0.001), while urban residence was modestly protective (HR = 0.98, 95% CI: 0.96–0.99, p = 0.005). Male sex (HR = 1.14, 95% CI: 1.13–1.16, p < 0.001) and older age (HR = 1.04, 95% CI: 1.04–1.04, p < 0.001) were also associated with greater hazard. Clinical predictors included lower baseline eGFR per −10 mL/min/1.73 m² (HR = 1.08, 95% CI: 1.07–1.08, p < 0.001), number of nephrology visits (HR = 1.07, 95% CI: 1.06–1.09, p < 0.001), and hospitalizations (HR = 1.37, 95% CI: 1.37–1.38, p < 0.001). Prescriptions for ACE inhibitors or ARBs, and PPIs were positively associated with MAKE (HR = 1.18, 95% CI: 1.16–1.19, p < 0.001; HR = 1.18, 95% CI: 1.16–1.19, p < 0.001, respectively), while NSAID use was associated with reduced hazard (HR = 0.91, 95% CI: 0.90–0.93, p < 0.001; Fig. 2b; online suppl. Table S7).

Similar to the diabetes cohort, the hypertension XGBoost model demonstrated better predictive performance than the Cox models, with significantly higher average precision (XGBoost 0.356; Cox 0.315, p < 0.001) and AUC-ROC (XGBoost 0.753; Cox 0.726, p < 0.001, online suppl. Table S2). Similar patterns were observed in the SHAP analysis of the hypertension cohort; older age, greater hospitalization counts, lack of commercial insurance, ACE inhibitor or ARB medication use, and lower baseline eGFR emerged as the top five influential features of the MAKE outcome. Similar to the diabetes cohort, these predictors were largely consistent with those identified in the Cox models, though differences in the relative importance were observed; Unknown and Medicaid insurance, urban residence, race, and nephrology care visits had little to no importance in the ML model (online suppl. Fig. S6, S7).

In the Cox proportional hazards subgroup analysis of the NHB-only population, insurance type remained the strongest predictor of MAKE. Medicaid was associated with the highest hazard (HR = 2.08, 95% CI: 1.88–2.30, p < 0.001), while risks for Medicare (HR = 1.74; 95% CI: 1.58–1.91, p < 0.001) and unclassified insurance (HR = 1.72; 95% CI: 1.56–1.89, p < 0.001) were also higher versus commercial insurance. Urban residence was protective (HR = 0.70; 95% CI: 0.61–0.80, p < 0.001), and the hazard of MAKE was higher among NHB men (reference: NHB women) (HR = 1.27, 95% CI: 1.19–1.36, p < 0.001) as in the total hypertension population comparing men to women (HR = 1.14, 95% CI: 1.13–1.16, p < 0.001; Fig. 3b; online suppl. Table S8).

Exploratory Analyses

Despite their clinical relevance, UACR and UPCR were missing for at minimum 88% of cases. After predictive mean matching imputation, their inclusion led to modest improvements in model performance and was most pronounced in the hypertension cohort where the average precision improved from 0.32 to 0.36 (online suppl. Table S9).

Discussion

This study investigated the relationship between race and ethnicity, social factors, and the risk of MAKE in NHB and NHW patients with diabetes and hypertension. In both cohorts, NHB patients had a significantly higher risk of MAKE compared to NHW patients after adjustment for demographic, social and clinical covariates. These findings support disproportionately higher risks of adverse kidney outcomes including loss of kidney function, kidney failure, and mortality among NHB patients with diabetes and hypertension in the US [1, 7, 9, 11, 15].

At cohort entry, NHB patients were younger than NHW patients, pointing to earlier and potentially greater exposure to hyperglycemia and hypertension during the lifespan, major drivers of CKD progression and mortality [27–31]. Despite relatively preserved kidney function at baseline (average eGFR of 83 mL/min/1.73 m² in the diabetes cohort and 87 mL/min/1.73 m² in the hypertension cohort), a substantial proportion of patients experienced ≥40% eGFR decline and all-cause mortality within 4 years. Risk of death has been shown to exceed that of kidney failure among patients with diabetes and higher eGFR [32]. These findings underscore the high risks of CKD progression in diabetic and hypertensive populations irrespective of race but when coupled with younger age at diagnosis for NHB patients, highlight a concerning trajectory of kidney function loss. For the hypertension population in particular, NHB patients were on average 9 years younger than NHW patients and had an apparent lower risk for MAKE as inferred from the unadjusted Kaplan-Meier analysis, yet the full model shows this relationship was reversed with adjustment for the variable set including age and social factors. This finding suggests that the younger age distribution among NHB patients may obscure the influence of social and structural risk, which become more apparent after accounting for differences in age and other covariates.

Beyond race and ethnicity, social factors were key predictors of MAKE. Insurance type was the strongest predictor in both diabetic and hypertensive cohorts, with Medicaid, Medicare, and unclassified insurance types conferring significantly higher risk than commercial insurance. While not a comprehensive measure of social risk, insurance status often reflects broader socioeconomic status and barriers to quality healthcare and preventative services [33–38]. Urban residence and SVI were also significant predictors. Urban living was protective for NHB patients, over 90% of whom lived in urban settings, consistent with evidence of greater access to transportation, healthcare facilities, and social support systems compared to NHB patients living in rural areas [39, 40]. Each tenth-of-a-unit increase in SVI was associated with higher MAKE risk, reinforcing evidence that cumulative social disadvantage contributes to adverse kidney outcomes, particularly among racially minoritized populations [35, 41–43]. Importantly, these findings point to the value of an intersectional perspective, as overlapping social and demographic identities such as race, sex, and neighborhood disadvantage may compound risk rather than operate independently [44–46].

Several clinical and healthcare utilization factors were shared predictors of MAKE in both the diabetes and hypertension cohorts even with adjustment for hypertension at baseline in the diabetes cohort, which was associated with an 8% higher risk for MAKE. Higher baseline eGFR was protective, underscoring the importance of preserving kidney function in mitigating adverse outcomes. Hospitalizations were associated with a 30–38% increased risk of a MAKE likely reflecting unmeasured disease severity or complications from uncontrolled hypertension or diabetes. Prescriptions for ACE inhibitors or ARBs, and PPIs were also associated with higher MAKE risk, likely due to treatment-by-indication bias [47], and perhaps ACE inhibitor and ARB induced physiological reduction of eGFR by up to 20–30% [48]. Some cohort-specific differences were also observed, further demonstrating intersectional patterns of risk. In the hypertension cohort, the risk of MAKE was most pronounced among NHB men compared to NHB women, aligning with well-documented evidence of disproportionate hypertension burden in the NHB male population (MEN???). Further, the complex interplay of biological, socioeconomic, and psychosocial risk factors, including stress, discrimination, and health behaviors, contribute to poor hypertension control and complications among NHB men in the US [49–57]. These findings underscore the continued need for intersectional interventions that address racial disparities within a social and structural context.

Our findings showed that ML models perform as well or better than Cox models in predicting survival outcomes [58–61]. The improved performance observed for the XGBoost model likely results from the ability of this approach to capture nonlinear interactions and handle missingness without imposing proportional hazards assumptions. SHAP values and feature ranking corroborated key predictors identified in the Cox models, including insurance status, hospitalizations, and baseline eGFR. Most notably, race was not among the most influential features. Despite well-documented racial disparities in kidney disease outcomes [11, 62, 63], the low SHAP importance ranking of race by XGBoost and small effect size by the Cox model suggests that race contributes little to the prediction of a MAKE outcome once other clinical and social factors are considered. Unlike Cox models, ML models allowed inclusion of UACR and UPCR by imputation despite high baseline missingness, modestly improving performance. While Cox proportional hazards models remain important for interpretable risk estimates, they require complete data and adherence to assumptions. By contrast, ML models handle non-linearities and missingness but often require additional explainability tools. A hybrid approach was most effective, leveraging ML models for hypothesis generation and prediction while relying on Cox proportional hazards models for risk quantification and clinical inferences.

Taken together, these findings suggest that social and structural factors are stronger predictors of MAKE than race alone. These findings point to a consistent and compounding influence of social factors, particularly insurance status, geographic context, and social vulnerability on kidney outcomes. While they do not fully account for the racial difference in MAKE outcomes, they represent actionable targets. Other observed differences in outcomes may reflect the contribution of additional factors not captured in this study, including environmental exposures and genetic risk factors such as apolipoprotein L1 (APOL1) variants, that increases the risk of kidney disease and are more common among individuals with West African ancestry [64]. Integrating social care into healthcare delivery through team-based models and risk-informed care pathways has been recommended by national organizations as a strategy to address these risks and promote fairness in chronic disease management [65]. Comprehensive assessment of patients’ social risks including unmeasured factors such as food instability, housing insecurity, and transportation barriers, may guide more effective interventions and improve long-term kidney outcomes. Additionally, very low rates of receiving UACR and UPCR tests, <13% of the diabetes cohort and <3% of the hypertension cohort, reflect poor adherence to guideline-directed practices [66–69]. Given the centrality of albuminuria/proteinuria for predicting adverse clinical outcomes [70–72], greater adherence to testing could substantially improve risk stratification and implementation of appropriate interventions.

This study has notable limitations. Race and ethnicity were assessed as recorded in the EHR and may not reflect self-identification. While prior research suggests strong concordance for NHB and NHW classifications [73–76], misclassification remains possible. Selection bias is also a concern, as analyses were limited to patients who accessed the health system. The single Western US health system may limit generalizability to other geographic regions and care settings. Further, the observed relationships between race, social factors, and MAKE may not directly generalize to non-US settings, where healthcare systems, social stratification, and the meaning of race differ. Future research is needed to evaluate whether similar patterns are observed in other countries and health systems. Although the inclusion of neighborhood-level social factors was a strength of this study, additional relevant individual-level social factors such as education level, employment status, housing and food stability and health literacy, were unavailable and may contribute to observed differences in MAKE risk. Additionally, UACR and UPCR data were missing for a large proportion of patients. Although, multiple imputation allowed inclusion of these variables and modestly improved model performance, the extent of missingness may introduce uncertainty into model performance and feature importance. Further, patients with dual insurance, though uncommon in this registry, were classified based on the highest-ranking insurance type, potentially biasing risk estimates. Baseline use of GLP-1 receptor agonist and SGLT2 inhibitor was less than 2% during the study period which predated widespread adoption of these medications and therefore, were not included in modeling. Clinical status and covariates were assessed at the baseline only, and the study methodology did not estimate risk associated with developing hypertension or diabetes during follow-up. The COVID-19 pandemic did not confound prediction of kidney failure in previous reports, although we cannot exclude such an effect in the present analyses [32]. Finally, we used consistent feature sets across modeling approaches for comparability. Future work should leverage the strengths of survival-aware ML (e.g., Random Survival Forest) with ML-specific feature engineering to model time-to-event outcomes directly, incorporate variables excluded here due to low prevalence [77], and explore the use of a competing risks framework to investigate predictors of kidney failure and death separately rather than as a composite outcome [32].

Conclusion

In large cohorts with diabetes or hypertension, NHB patients experienced a higher risk of MAKE compared to NHW patients, after accounting for demographic, social, and clinical factors. Insurance status, neighborhood-level social vulnerability and number of hospitalizations were consistent predictors across cohorts, underscoring the compounding role of structural and social risks in shaping kidney outcomes. Addressing modifiable social risks through integrated care models and improving guideline-directed testing offer actionable opportunities for earlier CKD identification, preserving kidney function, and improving overall and kidney health for people with diabetes or hypertension irrespective of race.

Acknowledgments

The authors acknowledge Providence Health Systems, the CURE-CKD Consortium, and the AIM-AHEAD Coordinating Center for their support of this work. The authors would also like to thank Drs. Lisa Cooper and Janice Bowie for their mentorship and thoughtful feedback on the dissertation that formed the foundation of this manuscript.

Statement of Ethics

This study was approved by Providence and Johns Hopkins University institutional review boards (IRB Approval No. STUDY2023001010 and 25753, respectively), which determined that written informed consent was not required for this limited data set.

Conflict of Interest Statement

K.N.M. received support from a federal research grant. A.Z. is supported by federal research grants. R.Z.A. is supported by federal research grants and reports having received prior support from the CDC, Travere Therapeutics Inc., the Doris Duke Charitable Foundation, The George Institute for Global Health, Benaroya Research Institute, as well as current support from Bayer AG, AstraZeneca, Novo Nordisk, CareDx, and KBP BioSciences, as well as personal fees from Boehringer Ingelheim, Bayer Pharmaceuticals, and Eli Lilly. L.M.K. is supported by a federal research grant, and by an investigator-initiated project funded by Otsuka Pharmaceutical, Inc. She also reports prior support from federal research grants, and other support from Benaroya Research Institute, and Travere Therapeutics Inc. C.R.J. is supported by an investigator-initiated project funded by Otsuka Pharmaceutical, Inc., and reports having received prior support from federal research grants, and other support from Benaroya Research Institute, Travere Therapeutics Inc, and the Doris Duke Charitable Foundation. K.B.D. is supported by a federal research grant, and by an investigator-initiated project funded by Otsuka Pharmaceutical, Inc. He also reports having received prior support from federal research grants, and other support from Benaroya Research Institute, Travere Therapeutics Inc, and the Doris Duke Charitable Foundation. C.L.R. received federal research grant support from the NIH and CDC, and other support from Travere Therapeutics Inc, and the Doris Duke Charitable Foundation. S.B.N. is supported by federal research grants and received research support from CDC. She reports personal fees and other support from Bayer AG and Boehringer Ingelheim, and reports speaker fees from Boehringer Ingelheim. R.J.T. is supported in part by federal research grants. P.P., and L.S. received support from a federal research grant. J.J.N. received support from a federal research grant and a research grant from the American College of Clinical Pharmacy. He reports consulting fees from Proteomics International, and serves on the Board of Directors for the American Diabetes Association. K.C.N. is supported in part by federal research grants and reports personal fees from Atlantis Dialysis Inc and AstraZeneca. K.R.T. reports support from federal research grants; investigator-initiated industry grant support from Otsuka, Bayer and Travere; non-profit grants from Breakthrough T1D, the Benaroya Research Institute, and the Doris Duke Charitable Foundation. She reports consultancy fees from Alnylam, AstraZeneca, Bayer, Boehringer Ingelheim, Glaxo Smith Kline, Lilly, Novo Nordisk, Otsuka, ProKidney, and Roche. She has received speaker fees from AstraZeneca, Bayer, Boehringer Ingelheim, Novo Nordisk, and Travere. Additionally, she reports being chair of data safety monitoring boards for the National Institute of Diabetes and Digestive and Kidney Disease, and George Clinical Institute, and member of the data safety monitoring board for AstraZeneca. She reports leadership roles as chair for the Diabetic Kidney Disease Collaborative chair for Kidney Week 2025 Program Committee both for the American Society of Nephrology, and a member of the American Heart Association/American College of Cardiology Cardiovascular-Kidney-Metabolic Guideline Committee, and the Diabetes and Chronic Kidney Disease Guideline Committee for Kidney Disease Improving Global Outcomes.

Funding Sources

This research was in part funded by the National Institutes of Health (NIH) Agreement No. 1OT2OD032581. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the NIH. The funder had no role in the design, data collection, data analysis, and reporting of this study.

Author Contributions

All authors met the International Committee of Medical Journal Editors criteria for authorship for the article. K.N.M., R.Z.A., and K.R.T. led project development from concept through data acquisition, curation, analyses, interpretation, and drafting the manuscript. A.Z., R.Z.A., L.M.K., C.R.J., K.B.D., C.L.R., R.J.T., P.P., L.S., S.B.N., K.C.N., and K.R.T. contributed to data acquisition, curation, analyses, and drafting the manuscript. K.N.M., L.M.K., P.P., and L.S. created the figures. J.J.N. reviewed and edited the manuscript. K.N.M., L.M.K., C.R.J., C.L.R., K.B.D., and K.R.T. accessed and verified the data.

Funding Statement

This research was in part funded by the National Institutes of Health (NIH) Agreement No. 1OT2OD032581. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the NIH. The funder had no role in the design, data collection, data analysis, and reporting of this study.

Data Availability Statement

Data that support the findings of this study are available upon reasonable request from the corresponding author. Data are not publicly available due to privacy or ethical restrictions and would require a data use agreement.