OBJECTIVE

End-stage kidney disease (ESKD) is a life-threatening complication of diabetes that can be prevented or delayed by intervention. Hence, early detection of people at increased risk is essential.

RESEARCH DESIGN AND METHODS

From a population-based cohort of 5,460 clinically diagnosed Danish adults with type 1 diabetes followed from 2001 to 2016, we developed a prediction model for ESKD accounting for the competing risk of death. Poisson regression analysis was used to estimate the model on the basis of information routinely collected from clinical examinations. The effect of including an extended set of predictors (lipids, alcohol intake, etc.) was further evaluated, and potential interactions identified in a survival tree analysis were tested. The final model was externally validated in 9,175 adults from Denmark and Scotland.

RESULTS

During a median follow-up of 10.4 years (interquartile limits 5.1; 14.7), 303 (5.5%) of the participants (mean [SD] age 42.3 [16.5] years) developed ESKD, and 764 (14.0%) died without having developed ESKD. The final ESKD prediction model included age, male sex, diabetes duration, estimated glomerular filtration rate, micro- and macroalbuminuria, systolic blood pressure, hemoglobin A1c, smoking, and previous cardiovascular disease. Discrimination was excellent for 5-year risk of an ESKD event, with a C-statistic of 0.888 (95% CI 0.849; 0.927) in the derivation cohort and confirmed at 0.865 (0.811; 0.919) and 0.961 (0.940; 0.981) in the external validation cohorts from Denmark and Scotland, respectively.

CONCLUSIONS

We have derived and validated a novel, high-performing ESKD prediction model for risk stratification in the adult type 1 diabetes population. This model may improve clinical decision making and potentially guide early intervention.

The observed incidence of end-stage kidney disease (ESKD) in people with type 1 diabetes has stabilized (1) or decreased over the past decades (24), probably in relation to the increased use of renin-angiotensin system (RAS) blockers. However, the decline in ESKD risk has been substantially lower compared with other common diabetes-related complications, such as cardiovascular disease (CVD) (1,5), and ESKD still remains a life-threatening complication (6) with a 10-fold increase in mortality rate in type 1 diabetes (3).

Fortunately, ESKD can be prevented or delayed by intensive glucose- and blood pressure–lowering therapy (7), and early detection is therefore essential. ESKD often develops in people with complicated and poorly controlled type 1 diabetes (6). This group also faces a high degree of pre-ESKD death, especially in older age (3). Because death precludes the occurrence of ESKD, a person’s risk of developing ESKD also depends on overall mortality risk. Not considering the “competing” risk of death is likely to overestimate the absolute risk of ESKD (8,9). Because the decision to initiate ESKD preventive treatment is often based on the absolute risk of developing ESKD, it is essential to estimate individual ESKD risk accurately.

Prediction models for ESKD in diabetes are scarce. Except for one study that used a composite outcome of end-stage renal failure, coronary heart disease, stroke, amputation, blindness, and death (10) and one study that predicted renal function decline (2), there are, to our knowledge, no ESKD risk models developed for the type 1 diabetes population. Three prediction models have been developed for cohorts of people with type 2 diabetes: one in New Zealand (11) and two in Chinese adults (12,13). Type 1 diabetes differs from type 2 diabetes in that people with type 1 diabetes are generally diagnosed at younger ages and therefore exposed to diabetes-related risk factors for ESKD, such as hyperglycemia and hypertension, for a longer time. Furthermore, while increased blood pressure, chronic kidney disease (CKD), and smoking appear to be risk factors for ESKD in both types of diabetes, obesity seems to play a larger role in type 2 diabetes (14), whereas age at diabetes diagnosis is mainly a risk factor in type 1 diabetes (3,14,15). This suggests a difference in the pathophysiology of ESKD for type 1 and type 2 diabetes, and prediction models specific to the type 1 diabetes population are needed.

Change in estimated glomerular filtration rate (eGFR) is a predictor of ESKD in diabetes (2), and the Kidney Disease Outcomes Quality Initiative clinical practice guidelines for diabetes and CKD suggest monitoring the rate of decline in eGFR to predict the time to onset of kidney failure (16). However, information on prior eGFR trajectory in people with type 1 diabetes requires continuous monitoring of eGFR, which is not widely feasible. Hence, the ability to assess ESKD risk in type 1 diabetes on the basis of current levels of risk factors is needed.

The aim of this study was to develop a risk prediction model for ESKD accounting for the competing risk of death, using a large population-representative cohort of adults with type 1 diabetes with an extensive range of clinical data and information on ESKD events and mortality from national registers. We externally validated the model in national and international cohorts to assess its broader generalizability.

Study Design and Participants

The study is based on a large population-based cohort of 5,506 adults with type 1 diabetes treated at the outpatient clinic at Steno Diabetes Center Copenhagen (SDCC) in the period from 1 January 2001 to 31 December 2016. In Denmark, treatment of people with type 1 diabetes is based in tertiary care, and referral to specialist care is free of charge. The type 1 diabetes population at SDCC includes the entire adult age span with both newly diagnosed and long-term diabetes, reflecting the background population with type 1 diabetes within this region. Individuals were followed from the date of their first clinical examination with a measurement of serum creatinine until first event of ESKD, death, emigration, or censor date of 31 December 2016 (date of register extraction).

To ensure exclusion of extreme values of metabolic risk factors, such as hemoglobin A1c (HbA1c) and lipids often present at the time of diagnosis, clinical examinations within the 1st year of diabetes diagnosis were excluded from the analyses. We further excluded people with prevalent ESKD at their first clinical examination (n = 46 [0.8%]), leaving 5,460 people with type 1 diabetes with a total of 42,921 clinical examinations for analysis.

According to Danish law, ethics approval and participant consent are not required for registry-based studies. Access and use of the described data were approved by the Danish Data Protection Agency (j-No. VD-2019-197) and the Danish Patient Safety Authority (j-No. 3-3013-2959/1).

Measurements and Definitions

Detailed clinical data of the participants were collected from the electronic health records at SDCC and linked to nationwide registries on mortality and morbidity, including ESKD (17,18), using the unique personal identification number given to all Danish residents at birth or at immigration (19). To separate type 1 from type 2 diabetes, type 1 diabetes was clinically diagnosed on the basis of phenotype and in accordance with the Danish National Diabetes Quality Database requirements (20). Participants were classified with type 1 diabetes if age at diagnosis was <30 years in combination with insulin treatment at diagnosis or if age at diagnosis was ≥30 years with randomly obtained nonfasting low C-peptide values (according to laboratory-specific reference values) or GAD65 antibody positivity, both in combination with a need for insulin to control blood glucose levels.

Electronic health data on all clinical visits with a measurement of serum creatinine were extracted together with the corresponding clinical and behavioral data. Detailed information on how measurements were obtained have been reported previously (21,22).

Albuminuria was classified from 24-h sterile urine collections (mg/24 h) or spot urine (mg/g) into normoalbuminuria (<30), microalbuminuria (30–299), or macroalbuminuria (≥300). We categorized smoking status into current smoking (yes/no), physical activity into regular physical activity defined as ≥30 min/day (yes/no), alcohol intake in three classes (0, 1–20, and >20 units/week), use of antihypertensive treatment (yes/no), lipid-lowering treatment (yes/no), and RAS-blocking treatment (yes/no). Retinopathy status was assessed from retinal photographs (no retinopathy, mild/moderate retinopathy, or severe retinopathy). eGFR was calculated using the Chronic Kidney Disease Epidemiology Collaboration equation (23).

Previous CVD was defined as any previous event of ischemic heart disease, ischemic stroke, heart failure, and peripheral artery disease as previously defined (22). We defined ESKD as a composite event of CKD stage 5 (ICD-10 code DN185), dialysis (procedure code BJFD), kidney transplantation (procedure code KKAS), or an eGFR <15 mL/min/1.73 m2. ESKD event data were obtained from the Danish National Patient Register (18). Data on date and cause of death were collected from the Cause of Death Register (17). Death without having developed ESKD was defined as non-ESKD–related mortality. Data on ethnicity were obtained from the Central Person Register (19), and ethnicity in the current study was defined as geographical region of origin (Europe, Middle East, or other). The registers are nationwide and cover all Danish residents.

Statistical Analysis

To account for the competing risk of death, cause-specific rate models for ESKD and death were estimated and then combined into a model for cumulative ESKD. We first developed a core model from commonly measured factors, including age, sex, diabetes duration, eGFR, albuminuria status, systolic and diastolic blood pressure, HbA1c, smoking, and previous CVD, and an extended model that further included RAS-blocking treatment, other antihypertensive treatment, lipid-lowering treatment, BMI, ethnicity, retinopathy, total and HDL cholesterol, LDL cholesterol, triglycerides, hemoglobin, alcohol intake, regular exercise, height, urinary albumin-to-creatinine ratio (UACR), potassium, sodium, and thyroid-stimulating hormone (TSH). Predefined interactions between clinical measurements and treatment, as well as other interactions between predictors identified in a prior conditional survival tree analysis, were included in both models. A survival tree analysis is a hypothesis-free data-driven approach and can be used to investigate potential high-level interactions between predictors for a specific outcome (24). In a subset of 4,815 (88%) participants with at least two clinical examinations, we further tested the effect of including eGFR annual change before baseline in the core model.

The cause-specific rate models for ESKD and death were estimated separately using Poisson models, with log of risk time as offset and censoring for the other event. For each participant, the follow-up period was split into 1-year age bands to account for the nonconstant effect of age over time on risk of ESKD and mortality (25) and then additionally split at the time points of repeated clinical measurements during follow-up. At each time interval, the most recent values of the predictors were used, and age and diabetes duration were updated. Before analysis, predictors with a highly skewed distribution were log2 transformed to improve model calibration. Backward elimination was used to test the predictors and interactions. The level of statistical significance was set at 5%. Postestimation shrinkage factors for the predictors were estimated in all the cause-specific rate models (26). Finally, the cumulative incidence of ESKD within a given time period for each person was calculated using the conditional survival function (8), which is based on both the estimated rate for ESKD and the estimated rate for death (see Supplementary Material for statistical details).

The discriminatory power of the models was evaluated using the C-statistic (27), with the CI computed from the DeLong method (28). In addition, model calibration was determined with the Hosmer-Lemeshow test of goodness of fit (29) by comparing means of estimated cumulative ESKD risk with the corresponding observed incidence in deciles of estimated risk.

For most covariates, <5% of the values were missing. However, for lipids, 13–16% of data were missing, and 26% were missing for TSH. To avoid exclusion of participants with missing values on the covariates, which may infer biased results (30), analyses where performed in 50 imputed data sets using multivariate imputations by chained equations method (31) with a missing-at-random assumption. Estimates of parameters of interest were summarized across the imputation copies according to Rubin rules (32).

Validation

The cumulative incidence functions for ESKD with the original regression coefficients in both the core and the extended models were internally validated using the first clinical examination of the derivation cohort. The cumulative incidence function for ESKD with both the original and the shrunken regression coefficients for the core model was externally validated nationally in the type 1 diabetes population of the Danish Funen Diabetes Database (FDDB) (33) and internationally in the Scottish Diabetes Research Network Type 1 Bioresource (SDRNT1BIO) (2). The validation cohorts did not have the required data available for validation of the extended model.

In FDDB, where we had access to baseline eGFR, discrimination and model calibrations for a 5- and 10-year ESKD event among participants with eGFR ≥60 mL/min/1.73 m2 at baseline was also calculated. This subgroup constituted 91% of the FDDB study participants but accounted for only 41% and 46% of the ESKD events after 5 and 10 years of follow-up, respectively. Statistical analyses were performed using R 3.6.1 software (R Foundation for Statistical Computing, https://www.r-project.org).

The derivation cohort was mainly (91%) of European origin. Baseline characteristics are given in Table 1. At baseline, 7% had CKD stage 3 or 4, and 37% developed ESKD during follow-up. The majority of the ESKD cases were among participants with an eGFR ≥60 mL/min/1.73 m2 (Supplementary Table 1). This group was characterized by a high degree of micro- or macroalbuminuria (45% vs. 19% in the total cohort). Participants were followed for a median of 10.4 years (interquartile limits 5.1; 14.7), during which 303 (5.5%) developed ESKD and 764 (14.0%) died as a result of non-ESKD–related causes without having developed ESKD. The incidence rate of ESKD was 5.7 per 1,000 person-years.

Table 1

Characteristics of the study populations at their first clinical examination in the derivation and validation cohorts

Derivation cohortValidation cohorts
SDCCFDDBSDRNT1BIO
Participants, n 5,460 3,150 6,025 
Follow-up time (years) 10.4 (5.1; 14.7) 10.7 (5.8; 13.6) 6.9 (6.2; 7.4) 
Region of origin    
 Europe 91.3 — — 
 Middle East 6.1 — — 
 Other 2.6 — — 
Age (years) 42.3 (16.5) 42.8 (16.7) 45.1 (15.0) 
Male 54.1 57.8 56.1 
Age at diabetes diagnosis (years) 21.0 (12.2; 33.8) 26.1 (14.3; 40.4) 21.6 (12.1; 32.2) 
Diabetes duration (years) 15.6 (6.6; 27.4) 10.9 (2.5; 21.9) 20.3 (11.0; 30.9) 
HbA1c (mmol/mol) 69.4 (16.7) 67.0 (18.8) 71.1 (17.0) 
HbA1c (%) 8.5 (1.5) 8.3 (1.7) 8.7 (2.0) 
BMI (kg/m224.7 (3.7) 25.2 (4.3) 27.1 (5.0) 
UACR (mg/g) 8.0 (4.0; 19.0) 10.6 (4.5; 26.8) 8.8 (5.1; 25.7) 
Albuminuria    
 Normal 81.0 76.8 83.1 
 Micro 13.0 19.9 12.6 
 Macro 6.0 3.4 4.3 
Serum creatinine (μmol/L) 82.0 (69.0; 93.0) — — 
eGFR (mL/min/1.73 m299.8 (84.3; 114.6) 89.5 (75.1; 104.2) 100.0 (84.3; 111.4) 
eGFR categories (mL/min/1.73 m2   
 eGFR ≥90 66.6 49.3 — 
 60 ≤ eGFR < 90 26.3 41.8 — 
 30 ≤ eGFR < 60 6.0 8.1 — 
 15 ≤ eGFR < 30 1.1 0.8 — 
Hemoglobin (mmol/L) 8.7 (0.8) — — 
Potassium (mmol/L) 4.0 (0.4) — — 
Sodium (mmol/L) 138.8 (3.0) — — 
TSH (×10−3 IU/L) 1.5 (0.9; 2.2) — — 
Systolic blood pressure (mmHg) 132.1 (19.2) 130.9 (18.3) 130.8 (16.0) 
Diastolic blood pressure (mmHg) 78.0 (10.0) 77.7 (10.5) 75.4 (10.0) 
Total cholesterol (mmol/L) 4.9 (1.0) 4.9 (1.0) — 
HDL cholesterol (mmol/L) 1.6 (0.5) 1.7 (0.5) 1.5 (0.5) 
LDL cholesterol (mmol/L) 2.7 (0.9) 2.7 (0.9) 2.5 (1.0) 
Triglycerides (mmol/L) 1.0 (0.7; 1.5) 1.0 (0.7; 1.5) 1.1 (0.8; 1.7) 
RAS blockers 21.6 8.9 37.3 
Other antihypertensive treatment 26.9 7.1 28.1 
Lipid-lowering medication 10.4 9.3 52.1 
Retinopathy status    
 None apparent 46.3 50.1 38.5 
 Mild/moderate 22.0 32.5 34.9 
 Severe 31.7 17.5 26.6 
Current smoking 51.3 27.1 21.9 
Alcohol intake#    
 0 units/week 14.5 — 15.2 
 1–20 units/week 80.6 — 71.9 
 >20 units/week 4.8 — 12.8 
Regular exercise 69.1 56.7 40.0 
Previous CVD 8.9 9.4 6.8 
Derivation cohortValidation cohorts
SDCCFDDBSDRNT1BIO
Participants, n 5,460 3,150 6,025 
Follow-up time (years) 10.4 (5.1; 14.7) 10.7 (5.8; 13.6) 6.9 (6.2; 7.4) 
Region of origin    
 Europe 91.3 — — 
 Middle East 6.1 — — 
 Other 2.6 — — 
Age (years) 42.3 (16.5) 42.8 (16.7) 45.1 (15.0) 
Male 54.1 57.8 56.1 
Age at diabetes diagnosis (years) 21.0 (12.2; 33.8) 26.1 (14.3; 40.4) 21.6 (12.1; 32.2) 
Diabetes duration (years) 15.6 (6.6; 27.4) 10.9 (2.5; 21.9) 20.3 (11.0; 30.9) 
HbA1c (mmol/mol) 69.4 (16.7) 67.0 (18.8) 71.1 (17.0) 
HbA1c (%) 8.5 (1.5) 8.3 (1.7) 8.7 (2.0) 
BMI (kg/m224.7 (3.7) 25.2 (4.3) 27.1 (5.0) 
UACR (mg/g) 8.0 (4.0; 19.0) 10.6 (4.5; 26.8) 8.8 (5.1; 25.7) 
Albuminuria    
 Normal 81.0 76.8 83.1 
 Micro 13.0 19.9 12.6 
 Macro 6.0 3.4 4.3 
Serum creatinine (μmol/L) 82.0 (69.0; 93.0) — — 
eGFR (mL/min/1.73 m299.8 (84.3; 114.6) 89.5 (75.1; 104.2) 100.0 (84.3; 111.4) 
eGFR categories (mL/min/1.73 m2   
 eGFR ≥90 66.6 49.3 — 
 60 ≤ eGFR < 90 26.3 41.8 — 
 30 ≤ eGFR < 60 6.0 8.1 — 
 15 ≤ eGFR < 30 1.1 0.8 — 
Hemoglobin (mmol/L) 8.7 (0.8) — — 
Potassium (mmol/L) 4.0 (0.4) — — 
Sodium (mmol/L) 138.8 (3.0) — — 
TSH (×10−3 IU/L) 1.5 (0.9; 2.2) — — 
Systolic blood pressure (mmHg) 132.1 (19.2) 130.9 (18.3) 130.8 (16.0) 
Diastolic blood pressure (mmHg) 78.0 (10.0) 77.7 (10.5) 75.4 (10.0) 
Total cholesterol (mmol/L) 4.9 (1.0) 4.9 (1.0) — 
HDL cholesterol (mmol/L) 1.6 (0.5) 1.7 (0.5) 1.5 (0.5) 
LDL cholesterol (mmol/L) 2.7 (0.9) 2.7 (0.9) 2.5 (1.0) 
Triglycerides (mmol/L) 1.0 (0.7; 1.5) 1.0 (0.7; 1.5) 1.1 (0.8; 1.7) 
RAS blockers 21.6 8.9 37.3 
Other antihypertensive treatment 26.9 7.1 28.1 
Lipid-lowering medication 10.4 9.3 52.1 
Retinopathy status    
 None apparent 46.3 50.1 38.5 
 Mild/moderate 22.0 32.5 34.9 
 Severe 31.7 17.5 26.6 
Current smoking 51.3 27.1 21.9 
Alcohol intake#    
 0 units/week 14.5 — 15.2 
 1–20 units/week 80.6 — 71.9 
 >20 units/week 4.8 — 12.8 
Regular exercise 69.1 56.7 40.0 
Previous CVD 8.9 9.4 6.8 

Data are mean (SD), median (interquartile limits), or % unless otherwise indicated.

#

One unit alcohol = 12 g of pure alcohol.

Regular exercise: ≥30 min/day.

The final core model for cumulative risk of ESKD included age, sex, diabetes duration, eGFR, micro- and macroalbuminuria, systolic blood pressure, HbA1c, smoking, and previous CVD. Older age was associated with a lower rate of ESKD but with a higher rate of mortality. For the remaining predictors, more unfavorable levels were associated with higher rates of both ESKD and death (Table 2). In the extended model, increasing levels of hemoglobin and mild/moderate retinopathy were associated with a higher rate of ESKD but a lower rate of death. Higher levels of UACR were further associated with a higher risk of ESKD and death. BMI, triglycerides, regular exercise, and sodium were associated with the rate of mortality and thereby indirectly associated with the cumulative risk of ESKD (Supplementary Table 2). Overall, the results of the survival tree analyses were consistent with the difference in baseline characteristics between individuals who did and did not develop ESKD (Supplementary Table 1 and Supplementary Figs. 14).

Table 2

Incidence rate ratios (IRR) for predictors of ESKD and death: core model

ESKDNon-ESKD death
IRR (95% CI)P valueIRR (95% CI)P value
Age (10 years) 0.83 (0.75; 0.92) <0.001 —  
Male sex (vs. female sex) 1.40 (1.11; 1.78) 0.005 5.47 (2.58; 11.61) <0.001 
Diabetes duration (10 years) 1.13 (1.02; 1.25) 0.022 1.11 (1.05; 1.16) <0.001 
eGFR (halving) 8.15 (6.88; 9.65) <0.001 1.28 (1.11; 1.49) <0.001 
Microalbuminuria (vs. normoalbuminuria) 1.09 (0.76; 1.55) 0.643 1.64 (1.39; 1.94) <0.001 
Macroalbuminuria (vs. normoalbuminuria) 1.89 (1.32; 2.70) <0.001 2.39 (1.88; 3.02) <0.001 
Systolic blood pressure (10 mmHg) 1.08 (1.03; 1.14) 0.004 0.90 (0.86; 0.93) <0.001 
HbA1c (10 mmol/mol) 1.12 (1.03; 1.20) 0.005 1.10 (1.04; 1.15) <0.001 
Smoking (vs. no smoking) 1.27 (1.00; 1.62) 0.048 1.88 (1.63; 2.18) <0.001 
Previous CVD event (vs. no) 1.35 (1.05; 1.74) 0.019 1.93 (1.65; 2.25) <0.001 
Age (10 years), women —  2.37 (2.18; 2.59) <0.001 
Age (10 years), men —  1.92 (1.78; 2.07) <0.001 
ESKDNon-ESKD death
IRR (95% CI)P valueIRR (95% CI)P value
Age (10 years) 0.83 (0.75; 0.92) <0.001 —  
Male sex (vs. female sex) 1.40 (1.11; 1.78) 0.005 5.47 (2.58; 11.61) <0.001 
Diabetes duration (10 years) 1.13 (1.02; 1.25) 0.022 1.11 (1.05; 1.16) <0.001 
eGFR (halving) 8.15 (6.88; 9.65) <0.001 1.28 (1.11; 1.49) <0.001 
Microalbuminuria (vs. normoalbuminuria) 1.09 (0.76; 1.55) 0.643 1.64 (1.39; 1.94) <0.001 
Macroalbuminuria (vs. normoalbuminuria) 1.89 (1.32; 2.70) <0.001 2.39 (1.88; 3.02) <0.001 
Systolic blood pressure (10 mmHg) 1.08 (1.03; 1.14) 0.004 0.90 (0.86; 0.93) <0.001 
HbA1c (10 mmol/mol) 1.12 (1.03; 1.20) 0.005 1.10 (1.04; 1.15) <0.001 
Smoking (vs. no smoking) 1.27 (1.00; 1.62) 0.048 1.88 (1.63; 2.18) <0.001 
Previous CVD event (vs. no) 1.35 (1.05; 1.74) 0.019 1.93 (1.65; 2.25) <0.001 
Age (10 years), women —  2.37 (2.18; 2.59) <0.001 
Age (10 years), men —  1.92 (1.78; 2.07) <0.001 

The estimated impact of calendar time was small (<2% difference in incidence rate per calendar year) and was not statistically significant (P ≥ 0.241 for ESKD, P ≥ 0.066 for death). Hence, calendar time was not included.

The core model showed excellent and robust discrimination, with C-statistics of ≥0.872 over the 10 years of follow-up in the derivation data. Model calibration was good for up to 5 years. The extended model had slightly better performance, with C-statistics of ≥0.883 and good calibration for up to 6 years (Fig. 1 and Supplementary Table 3). Details on the estimated model parameters with and without postestimation shrinkage and how to apply them in the cumulative risk model are given in Supplementary Tables 4 and 5.

Figure 1

C-statistics for an ESKD event within years of follow-up time in the derivation (A) and validation (C) cohorts and P values for test of adequate model fit in the derivation (B) and validation (D) cohorts. The dotted lines in B and C denote the threshold for acceptable model calibration (acceptable above the dotted line).

Figure 1

C-statistics for an ESKD event within years of follow-up time in the derivation (A) and validation (C) cohorts and P values for test of adequate model fit in the derivation (B) and validation (D) cohorts. The dotted lines in B and C denote the threshold for acceptable model calibration (acceptable above the dotted line).

Close modal

In the sensitivity analysis, prebaseline change in eGFR in the core model had little effect, with an incidence rate ratio <1% for a 10-unit difference in eGFR change (P = 0.078). Also, discrimination for a 5- and 10-year ESKD event was not improved (P ≥ 0.290), and model calibration was unchanged.

Validation

The Danish FDDB cohort of 3,150 adults was followed between 1 January 2003 and 31 December 2016 and was representative of the type 1 diabetes population in that region. The participants were, on average, 5 years older at diabetes diagnosis; macroalbuminuria and severe retinopathy were less frequent; and current smoking was around one-half of that in the derivation cohort (Table 1). Median (interquartile limits) years of follow-up were 10.7 (5.8; 13.6), during which 147 (4.7%) participants developed ESKD and 422 (13.3%) died as a result of non-ESKD causes, corresponding to an incidence rate of ESKD of 4.9 per 1,000 person-years. The core model without shrinkage of the parameters performed best. Discrimination was excellent and robust over time, with a C-statistic of 0.871 for an ESKD event within 5 years and 0.866 for an event within 10 years. Model calibration was good for up to 5–6 years of follow-up (Fig. 1 and Supplementary Table 3). In the subgroup with baseline eGFR ≥60 mL/min/1.73 m2, the C-statistics were 0.744 (95% CI 0.641; 0.847) and 0.775 (0.711; 0.840) for a 5- and 10-year ESKD event, respectively. Model calibration was adequate (P ≥ 0.097).

The SDRNT1BIO cohort of 6,025 adults was followed between 1 January 2011 and 31 December 2018 and was representative of the type 1 diabetes population in Scotland. The SDRNT1BIO population was slighter older and with ∼5 years longer diabetes duration at baseline. The majority (89%) was of White ethnicity. Like the FDDB population, this cohort had less macroalbuminuria and severe retinopathy, and the prevalence of current smoking was around one-half of that in the derivation cohort. The SDRNT1BIO population had almost twice as many using RAS-blocking agents and five times as many using lipid-lowering medications (Table 1). Median (interquartile limits) follow-up was 6.9 (6.2; 7.4) years, during which 95 (1.6%) participants developed ESKD and 321 (5.3%) died as a result of non-ESKD causes, corresponding to an incidence rate of ESKD of 2.4 per 1,000 person-years (Table 1). The performance of the core model was similar with and without shrinkage of the parameters. For the model without shrinkage, discrimination was excellent and robust, with a C-statistic of 0.961 for an ESKD event within 5 years and 0.952 for an event within 8 years (the maximum follow-up time). Calibration was only borderline acceptable in the first 4–5 years for the core model (Fig. 1 and Supplementary Table 3).

We have derived and validated a high-performing model for predicting individual risk of ESKD in the adult type 1 diabetes population on the basis of predictors routinely collected in the clinic. An extension of the model to include less frequently measured factors did not substantially improve prediction, suggesting that the more parsimonious core model, which is more feasible in a clinical setting, is preferable for assessing individual 5-year ESKD risk in people with type 1 diabetes.

The ESKD cumulative incidence rates in the Danish derivation and validation cohorts were twice that of the validation cohort from Scotland and considerably higher than previously reported in Sweden and Finland (3,34). The annual incidence rates in the derivation cohort only decreased slightly over the 2001–2016 follow-up period. The referral criteria for people with type 1 diabetes are comparable in Denmark and Scotland, and the selection criteria for the cohorts were similar, except the Scottish cohort did not include people diagnosed with type 1 diabetes after age 50 years. A possible explanation for the difference in ESKD risk could be the more aggressive treatment with RAS blockers and a lower prevalence of smoking in Scotland.

The predictors included in the core model have previously been found to be associated with ESKD (24,15). However, no studies have combined them into a model for predicting individual risk of ESKD in type 1 diabetes. Only one model for ESKD has been developed in a population-representative type 2 diabetes cohort of mainly White ethnicity (11). When applied to our type 1 diabetes population, we found discrimination to be adequate for a 5-year ESKD event (C-statistic 0.790 [95% CI 0.688; 0.893]), but calibration was poor (P < 0.001).

Male sex was associated with an increased risk of ESKD in our model, which is in line with previous studies (3,15). A study from New Zealand found male sex to be associated with a decreased risk of ESKD but not when including eGFR in the model (11). The finding in a Swedish study of no risk difference in men and women diagnosed before puberty (34) was not supported by the survival tree analysis in our study, where no interaction between sex and age and diabetes duration was found.

Although, ESKD is more frequent in old age (4,15), the rate of ESKD decreased with age in our model. This is likely due to a healthy survivor effect driven by the strong association between age and mortality. In other words, old age seems to be a protective factor in ESKD because older people die before they develop ESKD. This finding is in line with another competing risk analysis of ESKD risk in people with type 1 diabetes who have macroalbuminuria and CKD stages 1–3 (15).

RAS-blocking agents and other antihypertensive treatment did not improve the models, which may be because of the association being mediated through an improvement in albuminuria and blood pressure. A similar result was found in 1,000 people with type 1 diabetes followed for 25 years in the U.S. where the association of antihypertensive medication with ESKD was lost when eGFR was included in the model (4). The same was found in a Caucasian type 1 diabetes population with macroalbuminuria and CKD stages 1–3 (15).

Some studies have shown a decline in ESKD incidence over time (3,4). However, calendar time was not associated with ESKD in our model, indicating that any observed decline in ESKD over the years is reflected in the change in risk factor levels (35). This is supported by the Wisconsin Epidemiologic Study of Diabetic Retinopathy (WESDR) from the U.S. where an observed decline in incidence of ESKD over time was explained by improvements in glycemic and blood pressure control (4).

Our core model is adequate for assessing 5-year risk of ESKD, but predictions beyond this are questionable. Previous models have also been primarily assessed for 5-year risk of ESKD (11,13), although one model in type 2 diabetes was well calibrated up to 8 years of follow-up (12).

Strengths and Limitations

We had access to detailed data from repeated clinical examinations for the study participants, which allowed us to update the values of the predictors during follow-up to give a more correct estimate of the associations between the predictors and the event. In addition, missing data were imputed, thereby removing selection bias.

The derivation cohort was mainly of White ethnicity, which may explain why ethnicity was not predictive in the models, and further validation in populations of non-White ethnicity is needed. Our models were developed on the basis of data collected at a single clinical examination. Although recent studies have found historical measures of eGFR to improve prediction of future eGFR levels in type 1 diabetes (2) and ESKD in the general population (36), prebaseline change in eGFR in addition to baseline eGFR level did not improve prediction of future ESKD in our study.

In the future, prediction models for ESKD may also benefit from the inclusion of novel biomarkers or various omics data. However, such biomarkers, which are not used or collected routinely in clinical practice (37), have yet to prove predictive beyond that of clinical data (38).

Clinical Perspective

Although age-specific prevalence and incidence of ESKD have been stable since 2006 in Denmark (1), the actual number of people with type 1 diabetes who develop ESKD is increasing because of the general aging of the population. Mortality is still 70% higher in type 1 compared with type 2 diabetes (39), and quality measures of diabetes care in Denmark indicate a less aggressive approach to managing cardiovascular risk factors in type 1 diabetes (40). Early treatment could prevent or at least postpone the development of ESKD and thereby reduce treatment expenses and increase quality of life.

Our prediction model was developed for the entire range of eGFR not within the ESKD diagnostic range. Although people with CKD stage 3 and 4 are likely already managed as a high-risk group, not all will develop ESKD. In contrast, the majority of the ESKD events occur among people considered at low risk with a baseline eGFR ≥60 mL/min/1.73 m2. Our model also performed well in this subpopulation, and we believe that ESKD risk assessment is relevant at all levels of eGFR.

In conclusion, we have derived and validated a novel, high-performing ESKD prediction model for risk stratification in the adult type 1 diabetes population. This model may improve clinical decision making and potentially guide early intervention.

This article contains supplementary material online at https://doi.org/10.2337/figshare.13526414.

Funding. The Steno Diabetes Centers (Aarhus and Copenhagen) are partially funded by an unrestricted donation from the Novo Nordisk Foundation.

Duality of Interest. D.V. and G.S.A. own shares in Novo Nordisk A/S. F.P. reports having received research grants from AstraZeneca and lecture fees from AstraZeneca, MSD, Janssen, Eli Lilly, Boehringer Ingelheim, Novo Nordisk A/S, and Novartis as well as being a consultant/advisory board member for AstraZeneca, Bayer, Amgen, and MSD. P.R. has served as a consultant for AstraZeneca, Astellas, Bayer, Boehringer Ingelheim, Gilead, Merck, Mundipharma, Vifor, Sanofi, and Novo Nordisk A/S (all honoraria to his institution) and received research grants from AstraZeneca and Novo Nordisk A/S. M.E.J. has received research grants from AstraZeneca, Amgen, Sanofi, and Boehringer Ingelheim (investigator-initiated research). M.E.J. also owns shares in Novo Nordisk A/S. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. D.V. and G.S.A. analyzed the data. D.V., F.P., P.R., and M.E.J. conceived of the study concept and design. All authors took part in the interpretation of the results, commented on the manuscript, and had final responsibility for the decision to submit for publication. D.V. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of the study were presented at the virtual 56th European Association for the Study of Diabetes Annual Meeting, 21–25 September 2020.

1.
Steno Diabetes Center Copenhagen
.
Diabetes demography, cardiometabolic risk factors, and acute- and chronic diabetes complications in a nationwide diabetes registry in Denmark: a retrospective nationwide cohort study to investigate diabetes related complications and mortality rates over time
,
2020
.
2.
Colombo
M
,
McGurnaghan
SJ
,
Bell
S
, et al.;
Scottish Diabetes Research Network (SDRN) Type 1 Bioresource Investigators and the Scottish Renal Registry
.
Predicting renal disease progression in a large contemporary cohort with type 1 diabetes mellitus
.
Diabetologia
2020
;
63
:
636
647
3.
Helve
J
,
Sund
R
,
Arffman
M
, et al
.
Incidence of end-stage renal disease in patients with type 1 diabetes
.
Diabetes Care
2018
;
41
:
434
439
4.
Lecaire
TJ
,
Klein
BEK
,
Howard
KP
,
Lee
KE
,
Klein
R
.
Risk for end-stage renal disease over 25 years in the population-based WESDR cohort
.
Diabetes Care
2014
;
37
:
381
388
5.
Gregg
EW
,
Li
Y
,
Wang
J
, et al
.
Changes in diabetes-related complications in the United States, 1990-2010
.
N Engl J Med
2014
;
370
:
1514
1523
6.
Bjerg
L
,
Hulman
A
,
Carstensen
B
,
Charles
M
,
Witte
DR
,
Jørgensen
ME
.
Effect of duration and burden of microvascular complications on mortality rate in type 1 diabetes: an observational clinical cohort study
.
Diabetologia
2019
;
62
:
633
643
7.
The Diabetes Control and Complications Research Group
.
Effect of intensive therapy on the development and progression of diabetic nephropathy in the Diabetes Control and Complications Trial
.
Kidney Int
1995
;
47
:
1703
1720
8.
Andersen
PK
,
Geskus
RB
,
de Witte
T
,
Putter
H
.
Competing risks in epidemiology: possibilities and pitfalls
.
Int J Epidemiol
2012
;
41
:
861
870
9.
Berry
SD
,
Ngo
L
,
Samelson
EJ
,
Kiel
DP
.
Competing risk of death: an important consideration in studies of older adults
.
J Am Geriatr Soc
2010
;
58
:
783
787
10.
Soedamah-Muthu
SS
,
Vergouwe
Y
,
Costacou
T
, et al
.
Predicting major outcomes in type 1 diabetes: a model development and validation study
.
Diabetologia
2014
;
57
:
2304
2314
11.
Elley
CR
,
Robinson
T
,
Moyes
SA
, et al
.
Derivation and validation of a renal risk score for people with type 2 diabetes
.
Diabetes Care
2013
;
36
:
3113
3120
12.
Lin
CC
,
Li
CI
,
Liu
CS
, et al
.
Development and validation of a risk prediction model for end-stage renal disease in patients with type 2 diabetes
.
Sci Rep
2017
;
7
:
10177
13.
Wan
EYF
,
Fong
DYT
,
Fung
CSC
, et al
.
Prediction of new onset of end stage renal disease in Chinese patients with type 2 diabetes mellitus - a population-based retrospective cohort study
.
BMC Nephrol
2017
;
18
:
257
14.
Retnakaran
R
,
Cull
CA
,
Thorne
KI
,
Adler
AI
,
Holman
RR
.
Risk factors for renal dysfunction in type 2 Diabetes: U.K. Prospective Diabetes Study 74
.
2006
;
55
:
1832
1839
15.
Skupien
J
,
Smiles
AM
,
Valo
E
, et al
.
Variations in risk of end-stage renal disease and risk of mortality in an international study of patients with type 1 diabetes and advanced nephropathy
.
Diabetes Care
2019
;
42
:
93
101
16.
Kidney Disease: Improving Global Outcomes (KDIGO) Diabetes Work Group
.
KDIGO 2020 clinical practice guideline for diabetes management in chronic kidney disease
.
Kidney Int
2020
;
98
(
4s
):
S1
S115
17.
Helweg-Larsen
K
.
The Danish register of causes of death
.
Scand J Public Health
2011
;
39
(
Suppl.
):
26
29
18.
Lynge
E
,
Sandegaard
JL
,
Rebolj
M
.
The Danish National Patient Register
.
Scand J Public Health
2011
;
39
(
Suppl.
):
30
33
19.
Ebbesen
A
.
The creation of the central person registry in Denmark
. In
History of Nordic Computing 4: 4th IFIP WG 9.7 Conference, HiNC4, Copenhagen, Denmark, August 13–15, 2014, Revised Selected Papers (IFIP Advances in Information and Communication Technology (447))
.
Gram
C
,
Rasmussen
P
,
Østergaard
SD
, Eds. Cham, Switzerland,
Springer International Publishing
,
2015
, pp.
49
57
20.
Regionernes Kliniske Kvalitetsudviklingsprogram
.
Dansk Voksen Diabetes Database: datadefinitioner
,
2019
.
21.
Vistisen
D
,
Andersen
GS
,
Hulman
A
,
Persson
F
,
Rossing
P
,
Jørgensen
ME
.
Progressive decline in estimated glomerular filtration rate in patients with diabetes after moderate loss in kidney function-even without albuminuria
.
Diabetes Care
2019
;
42
:
1886
1894
22.
Vistisen
D
,
Andersen
GS
,
Hansen
CS
, et al
.
Prediction of first cardiovascular disease event in type 1 diabetes mellitus: the Steno Type 1 Risk Engine
.
Circulation
2016
;
133
:
1058
1066
23.
Levey
AS
,
Stevens
LA
,
Schmid
CH
, et al.;
CKD-EPI (Chronic Kidney Disease Epidemiology Collaboration)
.
A new equation to estimate glomerular filtration rate
.
Ann Intern Med
2009
;
150
:
604
612
24.
Hothorn
T
,
Hornik
K
,
Zeileis
A
.
Unbiased recursive partitioning: a conditional inference framework
.
J Comput Graph Stat
2006
;
15
:
651
674
25.
Carstensen
B
.
Age–period–cohort models for the Lexis diagram, Statistics in Medicine 2007; 26:3018–3045
.
Stat Med
2008
;
27
:
1561
1564
26.
Verweij
PJM
,
Van Houwelingen
HC
.
Cross-validation in survival analysis
.
Stat Med
1993
;
12
:
2305
2314
27.
Zweig
MH
,
Campbell
G
.
Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine
.
Clin Chem
1993
;
39
:
561
577
28.
Efron
B
,
Tibshirani
RJ
.
An Introduction to the Bootstrap
.
New York
,
Chapman & Hall
,
1993
29.
Gail
MH
,
Pfeiffer
RM
.
On criteria for evaluating models of absolute risk
.
Biostatistics
2005
;
6
:
227
239
30.
Janssen
KJ
,
Donders
AR
,
Harrell
FE
 Jr
., et al
.
Missing covariate data in medical research: to impute is better than to ignore
.
J Clin Epidemiol
2010
;
63
:
721
727
31.
van Buuren
S
.
Multiple imputation of discrete and continuous data by fully conditional specification
.
Stat Methods Med Res
2007
;
16
:
219
242
32.
Marshall
A
,
Altman
DG
,
Holder
RL
,
Royston
P
.
Combining estimates of interest in prognostic modelling studies after multiple imputation: current practice and guidelines
.
BMC Med Res Methodol
2009
;
9
:
57
33.
Adelborg
K
,
Szentkúti
P
,
Henriksen
JE
, et al
.
Cohort profile: the Funen Diabetes Database-a population-based cohort of patients with diabetes in Denmark
.
BMJ Open
2020
;
10
:
e035492
34.
Möllsten
A
,
Svensson
M
,
Waernbaum
I
, et al.;
Swedish Childhood Diabetes Study Group
;
Diabetes Incidence Study in Sweden
;
Swedish Renal Registry
.
Cumulative risk, age at onset, and sex-specific differences for developing end-stage renal disease in young patients with type 1 diabetes: a nationwide population-based cohort study
.
Diabetes
2010
;
59
:
1803
1808
35.
Andrésdóttir
G
,
Jensen
ML
,
Carstensen
B
, et al
.
Improved prognosis of diabetic nephropathy in type 1 diabetes
.
Kidney Int
2015
;
87
:
417
426
36.
van den Brand
JAJG
,
Dijkstra
TMH
,
Wetzels
J
, et al
.
Predicting kidney failure from longitudinal kidney function trajectory: a comparison of models
.
PLoS One
2019
;
14
:
e0216559
37.
Colhoun
HM
,
Marcovecchio
ML
.
Biomarkers of diabetic kidney disease
.
Diabetologia
2018
;
61
:
996
1011
38.
Colombo
M
,
Valo
E
,
McGurnaghan
SJ
, et al.;
FinnDiane Study Group and the Scottish Diabetes Research Network (SDRN) Type 1 Bioresource Collaboration
.
Biomarker panels associated with progression of renal disease in type 1 diabetes
.
Diabetologia
2019
;
62
:
1616
1627
39.
Carstensen
B
,
Rønn
PF
,
Jørgensen
ME
.
Prevalence, incidence and mortality of type 1 and type 2 diabetes in Denmark 1996-2016
.
BMJ Open Diabetes Res Care
2020
;
8
:
e001071
40.
Dansk Diabetes Database
.
National årsrapport 2017/2018
.
Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at https://www.diabetesjournals.org/content/license.