Predicting major outcomes in type 1 diabetes: a model development and validation study

Abstract

Aims/hypothesis

Type 1 diabetes is associated with a higher risk of major vascular complications and death. A reliable method that predicts these outcomes early in the disease process would be helpful in risk classification. We therefore developed such a prognostic model and quantified its performance in independent cohorts.

Methods

Data were analysed of 1,973 participants with type 1 diabetes who were followed for seven years in the EURODIAB Prospective Complications Study. Strong prognostic factors of major outcomes were combined in a Weibull regression model. The model performance was tested in three different prospective cohorts: Pittsburgh Epidemiology of Diabetes Complications study (EDC, n=554), Finnish Diabetic Nephropathy study (FinnDiane, n=2,999) and Coronary Artery Calcification in Type 1 Diabetes study (CACTI, n=580). Major outcomes included major coronary heart disease, stroke, end-stage renal failure, amputations, blindness and all-cause death.

Results

95 EURODIAB patients with type 1 diabetes developed major outcomes during follow-up. Prognostic factors were age, glycated haemoglobin, waist-hip ratio, albumin/creatinine ratio, and HDL cholesterol. A high risk group could be identified with 15% risk after 3-years of follow-up, 24% after 5-years and 32% after 7-years. The discriminative ability of the model was adequate with a C-statistic of 0.74. Discrimination was similar or even better in the independent cohorts: EDC, C-statistic = 0.79; FinnDiane, 0.82; and CACTI, 0.73.

Conclusions/Interpretation

Our prognostic model that uses easily accessible clinical features can discriminate between type 1 diabetes patients with good and poor prognosis. Such a prognostic model may be helpful in clinical practice and for risk stratification in clinical trials.

Introduction

Type 1 diabetes is a chronic autoimmune disease that can affect the cardiac and peripheral circulation. Many diabetes patients develop major complications related to the circulation later in life. Such major outcomes include coronary heart disease (CHD), stroke, end-stage renal failure, amputations and blindness[1–3], which decrease the quality of life considerably [4]. It has been found that patients with type 1 diabetes have a 4–8 fold increased CHD risk compared to those without diabetes [5, 6].

An important lesson learned from the Diabetes Control and Complications Trial (DCCT) was that intensive glycemic control aiming at near normal glycemic levels reduces the risk of all the diabetic complications including retinopathy, nephropathy and cardiovascular diseases [7–9]. Thus, intensive glycemic control should be initiated as early as possible in the course of diabetes, and not after 5 to 10 years when the first clinically detectable signs of complications such as retinopathy or nephropathy are already present [10].

Absolute risk predictions in individual patients with type 1 diabetes are important to timely identify the patients at high risk of major outcomes in order to enable strategies to prevent the development of such complications and to reduce health care costs. Further, prognostic models have an important role in informing the patient and to select high-risk populations for randomized controlled trials (RCTs).

Prognostic research in type 1 diabetes has so far focussed on CHD as a major outcome, disregarding stroke, end-stage renal failure, amputations and blindness [11, 12]. However, from a patient perspective, it is relevant to predict all the major outcomes, since any of these outcomes will lead to restricted quality of life in the relatively young patients. More importantly the presence of nephropathy is also associated with a substantially increased risk of premature death [13].

We here describe a prognostic model based on easily accessible patient and clinical characteristics for the composite of the major outcomes and death, which would allow straightforward application of the model in practice. This model may help to identify patients with type 1 diabetes at high risk of major outcome events. We use the large prospective cohort from the European Diabetes Prospective Complications Study for model development. The model performance was then tested in three different cohorts from North America and Europe.

Methods

Patients with type 1 diabetes

Data from the European Diabetes Prospective Complications Study (EURODIAB PCS) were used to develop the prognostic model (development set). EURODIAB PCS was designed to study risk factors for complications in patients with type 1 diabetes. The study included these patients from 31 centres in 16 European countries (see list of investigators in Online Supplementary Index 1). Ethics committee approval was obtained at each center, and all subjects provided written informed consent. Inclusion and predictor measurements occurred between 1989 and 1991. Full details of this study have been described previously [1, 14]. The major outcomes examined for the prognostic model were the following; major CHD, stroke, end-stage renal failure, amputation, blindness and death. We selected 1,973 study participants with no such complications at inclusion (between 1989 and 1991), and who were followed up to 1997–1999. The first occurrence of any of the major outcomes was used for analysis.

We used three other prospective cohorts including patients with type 1 diabetes to test the performance of the model: the Epidemiology of Diabetes Complications (EDC) study, the Finnish Diabetic Nephropathy (FinnDiane) study and the Coronary Artery Calcification in Type 1 Diabetes (CACTI) study (Table 1). The EDC study is a prospective investigation of factors that lead to the development of complications among individuals with childhood onset type 1 diabetes diagnosed or seen within one year of diagnosis at the Children’s Hospital of Pittsburgh, USA [15]. Participant inclusion occurred between 1986 and 1988. Since participants were re-examined biennially, two groups could be analysed: i) participants without severe comorbidity at baseline (EDC original, n = 554 ); ii) participants without severe comorbidity at a later examination (1996–1998, EDC recent, n = 314). The second selection allowed evaluation of the prognostic model in a group of more recently treated participants. Note that the participants from the EDC recent cohort are also member of the EDC original cohort. FinnDiane is an ongoing, prospective, nationwide, multicenter study with the aim to identify genetic and clinical risk factors for diabetic nephropathy in patients with type 1 diabetes [16]. Participants were included between 1994 and 2009. The CACTI study is a prospective cohort study of individuals with type 1 diabetes without baseline CHD designed to assess risk factors associated with the development and progression of subclinical CHD [17]. Participants were enrolled between 2000 and 2002. All three studies are ongoing. For the present analysis, we selected participants who had no major outcomes at inclusion. Follow-up measurements until 7 to 9 years after inclusion were taken to assess the occurrence of major outcomes.

Table 1.

Characteristics of patients with type 1 diabetes in the development and validation sets.

Characteristic	EURODIAB PCS 1989 – ‘91 (N= 1,973)	EDC original 1988 – ‘90 (N = 554)	EDC recent 1996 – ‘98 (N = 324)	FinnDiane 1994–2009 (N= 2,999)	CACTI 2000 – ‘02 (N = 580)
Age, years	30.3 (8.3)	28.0 (7.7)	36.2 (7.7)	37.3 (12.0)	36.3 (9.2)
Male gendera	1021 (52)	273 (49)	151 (48)	1521 (51)	261 (45)
Diabetes duration, yearsb	11.5 (1 – 26)	18.6 (8–38)	26.4 (18 – 46)	19.0 (1–55)	22.7 (4–52)
HbA1c, % [mmol/mol]	8.3 (1.9) [67]	9.1 (1.5) [76]	8.4 (1.4) [68]	8.4 (1.5) [68]	8.0 (1.3) [64]
Waist-hip ratio
Men	0.88 (0.08)	0.86 (0.05)	0.91 (0.05)	0.91 (0.07)	0.87 (0.06)
Women	0.80 (0.10)	0.77 (0.06)	0.81 (0.08)	0.81 (0.06)	0.78 (0.07)
Body mass index, kg/m2
Men	23.5 (2.7)	23.9 (2.9)	25.4 (3.1)	25.2 (3.3)	26.5 (3.7)
Women	23.3 (3.0)	23.7 (3.3)	25.1 (4.1)	25.0 (3.8)	26.0 (4.7)
Smoked
Ever	946 (48)	149 (35)	108 (34)	1242 (45)	111 (19)
Current	616 (31)	113 (20)	53 (17)	713 (26)	68 (12)
Systolic pressure, mmHg	118 (16)	113 (15)	115 (15)	132 (17)	117 (13)
Diastolic pressure, mmHg	75 (11)	73 (10)	70 (10)	79 (10)	77 (8)
Antihypertensive medication	106 (5)	45 (8)	55 (18)	959 (32)	198 (34)
Albumin creatinine ratio, mg/mmolb	1.2 (0.1 – 415)	2.4 (0.2 – 1086)	2.0 (0.1 – 807)	0.8 (0.0–853)	1.7 (1.1 – 547)
Fasting triglyceride, mmol/lb	0.9 (0.2 – 12.6)	0.9 (0.2 – 6.9)	1.0 (0.3 – 4.8)	1.0 (0.3–10.2)	0.9 (0.3–4.8)
HDL, mmol/l	1.5 (0.4)	1.4 (0.3)	1.4 (0.4)	1.3 (0.4)	1.5 (0.4)
Non-HDL, mmol/l	3.7 (1.1)	3.5 (1.1)	3.5 (0.9)	3.6 (1.0)	3.1 (0.8)
LDL, mmol/l	3.2 (0.9)	3.0 (0.9)	3.0 (0.8)	3.1 (0.9)	2.6 (0.7)
First incident Major outcomes	95 (5)	98 (18)	84 (26)	315 (11)	42 (7)
Death	33	11	8	61	15
CHD/Stroke	39	35	36	161	19
End-stage renal failure/Dialysis	0	16	11	88	1 (transplant)
Amputation	11	7	10	5	2
Blind	12	29	18	NA	5
Follow-up time, yearsb	7.4 (0.1 – 8.2)	8.1 (0.2 – 10.3)	8.1 (0.2–9.9)	7.5 (0.1–7.5)	7.3 (0.2 – 10.4)

Values are the mean (SD) unless otherwise noted.

^anumber (%) of patients;

^bmedian (range)

Major Outcomes

Major outcomes included major CHD, stroke, end-stage renal failure, amputations, blindness and all-cause death. Major CHD comprised fatal CHD, non-fatal myocardial infarction (MI) (coded according to ICD-9 410-414), and major Q waves on Minnesota coded ECGs, (codes 1.1, 1.2) [18]. Stroke (ICD-9 430-438) included fatal and non-fatal stroke. End-stage renal failure comprised renal dialysis or renal transplantation from hospital records. Amputation was ascertained by self-report and/or physician examination. Blindness was ascertained by self-report of physician-diagnosed blindness and by visual acuity testing using the Snellen chart [19]. Death was collected from death certificates. Time of occurrence of the major outcomes was assessed as precise as possible from hospital records and death certificates. Assessments of the major outcomes were similar for the three validation cohorts [15–17].

Prognostic factors for model development

Based on the literature and clinical expertise, the following characteristics were considered potential prognostic factors for the defined outcomes: age, gender, diabetes duration, glycated haemoglobin (HbA_1c), waist hip ratio (WHR), body mass index (BMI), smoking, systolic and diastolic blood pressure (BP), antihypertensive medication, albumin/creatinine ratio, fasting triglyceride, HDL cholesterol, non-HDL cholesterol and LDL cholesterol.

Demographic variables, diabetes duration, medication and smoking status were collected with standard questionnaires in the EURODIAB PCS study. Height, weight and waist and hip circumferences were measured in a standardized way to calculate BMI and WHR. Systolic and diastolic blood pressure was measured by a random zero sphygmomanometer and the mean of two measurements was calculated. Two 24 hour urine samples were analysed to determine mean urinary albumin and creatinine concentrations; albumin creatinine ratios were calculated. Fasting blood samples were taken to measure HbA_1c, and lipids. HbA_1cwas measured with an enzyme immunoassay using a monoclonal antibody. The HbA_1c values obtained were converted to DCCT values [20] and International Federation of Clinical Chemistry and Laboratory Medicine (mmol/mol) values in line with the European Association for the Study of Diabetes recommendations [21]. Lipids (fasting triglycerides, cholesterol and HDL) were measured with standard enzymatic methods. Non-HDL cholesterol was defined as the difference between total and HDL cholesterol and representing cholesterol carried on all of the potentially proatherogenic apoB-containing particles. LDL cholesterol was calculated from Friedewald's formula if triglycerides were <4 mmol/l [22]. Measurements of prognostic factors in the three validation cohorts were performed with similar methods [15–17].

Statistical analysis

Missing values occurred for most prognostic factors, varying from 2 to 92 missing values among 1,973 EURODIAB PCS participants. For the variables, LDL and triglycerides, more values (n=673) were missing. We imputed missing values with multiple imputation techniques using predictive mean matching, allowing all observed values to be analysed (aregImpute algorithm in R software) [23].

To develop the prognostic model, we used Weibull regression analysis to estimate univariate and multivariable regression coefficients, and hazard ratios with 95% confidence intervals for each prognostic factor. The functional form between continuous prognostic factors and occurrence of major outcomes was explored with restricted cubic splines.

A full multivariable model was fitted that included all candidate prognostic factors with chosen transformations. The number of prognostic factors was reduced with backward stepwise selection. Variables with a weak association (p>0.3) were deleted from the model. This analytical strategy aims to limit overfitting of a model to the available data [24]. Therefore, the backward selection procedure uses a liberal p value (0.3 in this study), which results in inclusion of relatively weaker prognostic factors in the model at the cost of possible selection of a variable without predictive value. Such a model can perform well in new participants [25]. We did not explore multiplicity (interaction terms) because of the relatively small sample size.

Internal validity was studied in 100 bootstrap samples to assess possible optimism. The regression coefficients in the final model were multiplied with a shrinkage factor, which was estimated with the bootstrapping procedure [24]. Shrinking the regression coefficients to zero reduces overconfidence in predicted probabilities.

External validity was assessed in: EDC (original and recent), FinnDiane and CACTI. Follow-up in the first EDC selection was truncated in order to analyse discrimination for a median follow-up time of 8 years, which is similar to the follow-up time of the development set. Discrimination was assessed with the Harrell’s C-statistic [26]. Discrimination was assessed graphically using Kaplan-Meier plots for three risk groups (high, intermediate, low risk). Calibration plots were made to compare observed to predicted risks at external validation.

Finally, a score chart was made based on the regression coefficients in the final model. Scores were calculated by the products of regression coefficients and prognostic factor values and rounded to integers. The sum scores were then related to 3-, 5- and 7-year risks of major outcomes.

Results

Development of the prognostic model

Major outcomes occurred within 7 years of follow-up in 95 of 1,973 participants in the EURODIAB PCS. Participants had a mean age of 30 years (SD 8.3) and the median duration of diabetes was 11.5 years (Table 1).

Table 2 shows the univariate associations of the possible prognostic factors and first incident major outcomes. All variables were positively associated with incident major outcomes, except for HDL cholesterol which showed an inverse association. The final prognostic model included age, glycated haemoglobin, waist-hip ratio, albumin/creatinine ratio, and HDL cholesterol (Table 3). Albumin/creatinine ratio was log transformed; all other continuous prognostic factors were included linearly. The strongest were age (Wald statistic=5.3) and albumin/creatinine ratio (Wald=4.4). The regression coefficients in the final model were multiplied with the estimated shrinkage factor of 0.93 (See Online Supplementary Index 2 for final regression formula). The C-statistic was 0.74 at internal validation.

Table 2.

Occurrence of severe complication for candidate prognostic factors in EURODIAB PCS

Patient characteristic	7-year risk of severe complicationa	Hazard ratio (95% CI)
Age, years
< 25	3%	1.0
25 – 34	4%	1.3 (0.7 – 2.5)
35 – 44	7%	2.7 (1.5 – 5.0)
45 +	15%	5.7 (2.9 – 11)
Gender
Female	4%	1.0
Male	6%	1.5 (1.0 – 2.3)
Diabetes duration, years
< 5	4%	1.0
5 – 14	4%	1.1 (0.6 – 2.0)
15 +	6%	1.6 (0.9 – 3.0)
HbA1c, % [mmol/mol]
< 7 [53]	3%	1.0
7 – 9.9 [53–85]	4%	1.5 (0.8 – 2.6)
10 + [85+]	8%	2.7 (1.5 – 5.0)
Body mass index, kg/m2
< 25	4%	1.0
25+	6%	1.4 (0.9 – 2.2)
Waist-hip ratio
< 0.8	2%	1.0
0.8 – 1.0	6%	2.4 (1.4 – 4.0)
1.0 +	10%	4.2 (1.8 – 9.6)
Ever smoked
Yes	6%	1.5 (1.0 – 2.3)
No	4%	1.0
Systolic pressure, mmHg
< 110	3%	1.0
110 – 130	4%	1.5 (0.9 – 2.7)
130 +	8%	2.8 (1.5 – 4.9)
Diastolic pressure, mmHg
< 80	4%	1.0
80 – 90	6%	1.5 (0.9 – 2.4)
90 +	8%	1.9 (1.1 – 3.4)
Antihypertensive medication
Yes	10%	2.3 (1.2 – 4.4)
No	4%	1.0
Albumin creatinine ratio, mg/mmol
< 0.5	4%	1.0
0.5 – 1.4	3%	0.9 (0.4 – 1.9)
1.5 +	7%	2.0 (0.9 – 4.5)
Fasting triglyceride, mmol/l
< 1	4%	1.0
1+	6%	1.7 (1.1 – 2.8)
HDL, mmol/l
< 1.2	6%	1.0
1.2 – 1.5	5%	0.9 (0.5 – 1.5)
1.5 +	4%	0.6 (0.4 – 1.1)
Non-HDL, mmol/l
< 3	2%	1.0
3 – 5	5%	2.3 (1.2 – 4.4)
5 +	9%	4.3 (2.1 – 8.9)

^aObserved risk assessed with Kaplan Meier curves

Table 3.

Multivariable associations of the selected prognostic factors with the occurrence of major outcomes.

Prognostic factors	Weibull regression coefficienta	HR	(95% CI)	P value
Age, per decade	− 0.661	1.93	(1.79 – 2.09)	< 0.001
HbA1c, %	− 0.161	1.17	(1.06 – 1.29)	0.001
Waist-hip ratio b	− 0.259	1.30	(1.09 – 1.54)	0.003
Albumin/creatinine ratio, mg/mmol c	− 0.310	1.36	(1.18 – 1.57)	< 0.001
HDL-c, mmol/l	0.306	0.74	(0.46 – 1.17)	0.199

^aWeibull regression results in opposite sign (− vs +) compared to Cox Proportional Hazards models , therefore HRs are estimated by exp(-Weibull regression coefficient). Estimates are for a unit increment, unless stated otherwise.

^b0.1 unit increase in waist-hip ratio (waist-hip ratio/10)

^clogarithmic increase in albumin/creatinine ratio (ln(albumin/creatinine ratio +1))

Figure 1 shows a score chart to facilitate the calculation of the absolute risk of major outcomes for an individual patient with type 1 diabetes. Values for continuous prognostic factors are given with small steps. A sum score for an individual patient consists of the sum of all scores. The lower part of Figure 1 shows the predicted 3-, 5- and 7-year risks that correspond to the sum scores (See Online Supplementary Index 3 for Excel format to automatically calculate the risk of major outcomes in patients with type 1 diabetes).

Figure 1.

Score chart to predict 3 -, 5 -, and 7 year absolute risk of major outcomes in type 1 diabetes.

Figure 1 A. Prognostic Model to predict 3, 5 and 7 year risk of major outcomes in patients with type 1 diabetes

Figure 1 B. Graph with 3, 5 and 7 year risk of major outcomes in patients with type 1 diabetes

We found that many participants had intermediate scores, 16 to 20. We therefore distinguished three (absolute) risk groups that were based on the sum scores: low risk with a score of 15 or less; intermediate with a score between 16 and 20; high with a score of 21 or more. The risk groups showed a 7-year risk of 1%, 4% and 11%. The high risk group contained 457 (23%) participants (Figure 2, upper left panel). The highest two risk groups (sum scores 16+) contained together 94% (n= 50+39 = 89) of all participants who developed major outcomes.

Figure 2.

Observed Kaplan Meier risk of major outcomes divided into three score groups

Kaplan-Meier estimates for the risk of major outcomes. Patients are categorised based on de total score (see Figure 1):

——— 8 – 15, low risk; 16 – 20, intermediate risk; 21+, high risk.

External validation of the prognostic model

Most participant characteristics from the external validation studies were comparable with the EURODIAB participants (Table 1). The mean age for EURODIAB and the original EDC cohort were similar with means of 30 and 28. The recent EDC had an 8 years older population than in the original EDC as the predictor values were measured 8 years later. FinnDiane and CACTI included older participants than EURODIAB. The median duration of diabetes differed between the studies from 11.5 years in EURODIAB to 26.4 in the EDC recent. Occurrence of major outcomes within 7 years was relatively high in the EDC original (14%), EDC recent (20%) and FinnDiane (11%) compared with the EURODIAB (4%) and CACTI (7%). The composition of major outcomes varied between the various cohorts.

The discriminative ability of the prognostic model was adequate in all cohorts with C-statistic of 0.79 in EDC, 0.74 in EDC recent and 0.73 in CACTI compared to a C-statistic of 0.74 in the development set. The discriminative ability tended to be even higher in the FinnDiane cohort than in the other cohorts with a C-statistic of 0.82. The relatively high discriminative ability in the FinnDiane was explained by more heterogeneity in prognostic factor values. The standard deviation of the linear prognostic factor (sumproduct of risk estimates and regression coefficients of the model) was 1.01 for FinnDiane compared with 0.81 for EURODIAB. The majority of participants with major outcomes were categorised in one of the two groups with the highest risk: 92/98 (94%) for EDC, 82/84 (98%) EDC recent, 306/315 (98%) for FinnDiane and 40/42 (95%) for CACTI.

The observed risks were much higher for the original and recent EDC participants than for participants from the other cohorts. Kaplan Meier estimates at 7 year were 15.9 (EDC), 21.5% (EDC recent), 10.3% (FinnDiane) and 6.9% (CACTI). As a consequence, the mean predicted risks were too low for the EDC cohorts (5.2% and 7.2%) and for FinnDiane (7.1%). The same pattern is shown for the three risk score groups. For instance, the risk at 7 years was in the EURODIAB participants 1% (low risk), 4% (medium risk) and 11% (high risk), with 5%, 7% and 34% for the original EDC and 7%, 10% and 34% for the recent EDC. High risk participants from FinnDiane had also higher observed risks than predicted (21% observed versus 15% predicted at 7 years) (Figure 2, to obtain risk estimates read y-axis for each Kaplan Meier curve at the 7 year point on the x-axis).

The underprediction is further shown in the calibration plots (Figure 3 left panels). Quintiles of risk estimates corresponded to the Kaplan-Meier estimates only for CACTI. To provide well calibrated risk estimates, a small model adjustment was necessary for participants from EDC and FinnDiane (Figure 3 right panels). The intercept of the Weibull model was adjusted for the EDC recent population to an intercept of 9.1 rather than 10.2; for participants from FinnDiane the intercept became 9.8.

Figure 3.

Calibration plots for observed and predicted risk at external validation for EDC, EDC recent, FinnDiane and CACTI. Triangles indicate the observed frequencies by quintile of predicted risks. Calibration of the model (left panels) was poor for EDC and FinnDiane. Adjustment of the model intercept (right panels) improved evidently the agreement between observed and predicted risks.

Discussion

This paper describes the development of a prognostic model for the risk of major outcomes (CHD, stroke, end-stage renal failure, amputations, blindness and death) in participants with type 1 diabetes. The prognostic model is based on common and easily measured prognostic factors which will allow for translation to clinical care. The model showed good discriminative ability in external validation cohorts. We found a systematic difference between predicted and observed complications in participants from the EDC and FinnDiane populations. The miscalibration disappeared after a small adjustment of the model.

Our prognostic model included: age, HbA_1c level, waist-hip ratio, albumin/creatinine ratio, and HDL cholesterol level. Most of these are known to be associated with CHD and other complications from etiological research in type 1 diabetes patients. HbA_1c level and albumin/creatinine ratio were shown to be strong prognostic factors for CHD, renal failure and death in type 1 diabetes [27]. A number of prognostic models for the major outcome CHD have been evaluated [28–32], only a few specifically for type 1 diabetes [12]. A prognostic model for CHD in participants with type 1 diabetes from the EDC study [12] included HDL cholesterol, micro- or macroalbuminuria, diabetes duration, and white blood cell count as the strongest prognostic factors in men and non-HDL cholesterol, waist-hip ratio, diabetes duration, systolic blood pressure, and use of antihypertensive medication in women. In our prognostic model, age was selected rather than diabetes duration. The high correlation between age and diabetes duration in EDC (Pearson correlation = 0.85) indicates that the two variables can include to a great extent the same information. This correlation was lower in the EURODIAB (0.43) with age being the stronger prognostic factor.

Other known prognostic factors, such as blood pressure or use of antihypertensive medication, LDL and smoking were not included in our model because of weak additional effects. Lacking some of these prognostic factors (blood pressure, LDL, smoking, family history) in our model does not mean these should not be measured or controlled. It implicates that these factors are less relevant for the prediction of major outcomes, particularly when other prognostic factors are already being considered. For instance, hypertension was correlated with the age and the albumin excretion rate. Older participants with higher albumin excretion rate have a relatively high blood pressure. With age and albumin excretion rate already in the model the added value of hypertension or blood pressure was limited. In the EURODIAB, we could study the inflammatory marker fibrinogen but the added predictive value was negligible (C-statistics for models with and without fibrinogen were 0.74). The minor additive value of predicting cardiovascular diseases with inflammatory markers such as fibrinogen and C-reactive protein was also recently demonstrated by the Emerging Risk Factors collaboration [33]. We did not examine sex-specific prognostic models since the number of participants with major outcomes in the development set was relatively low and would result in unstable risk estimates.

Our development cohort included participants from 16 European countries. The relatively heterogeneous cohort allowed the development of a model that generalised well to participants from other cohorts from Northern America and Europe. Particularly the cohorts FinnDiane and CACTI allowed validation in more recently treated patients with type 1 diabetes. Participants from the EDC and FinnDiane cohorts experienced more often major outcomes within 7 years than the participants from EURODIAB and CACTI. This could not completely be explained by higher values for the prognostic factors in the model. Hence the risk estimates may require regular updating according to specific population characteristics, such as calendar year or local organization of diabetes management [34, 35]. For the EDC and FinnDiane population, a simple intercept adjustment suffices.

Predicting major outcomes permits establishment of a risk profile for individual patients with type 1 diabetes. Physicians may consider active intervention in the identified high risk patients. Such interventions can include intensifying insulin regimen and cardiovascular risk management. We stress that our prognostic model is particularly able to identify patients with type 1 diabetes at the highest risk of complications. Clinicians should nevertheless still focus on aggressive lowering of all known prognostic factors. Further, physicians can use the risk assessments to inform the patient on lifestyle and diet. Prior research indicated benefits of diet and physical activity even in intensively treated patients with type 1 diabetes [36, 37]. The risk assessments also permits more efficient design and analysis of randomised controlled trials on the effectiveness of new interventions in patients with type 1 diabetes [38]. Patients with a very good or a very poor prognosis may be excluded; we can perform covariate adjustment of an intervention effect to increase statistical power [39].

Our study has several limitations. Relatively low numbers of participants had major outcomes, as most participants were young. Using the composite of CHD, stroke, renal failure, amputations, blindness and death, more participants experienced one of the outcomes, which increased the statistical power [40]. The trade-off of using a composite endpoint is that predictor effects of different outcomes are averaged [40]. The outcomes of diabetes that we combined are related to the circulatory system, essentially vascular. It has been shown that microvascular outcomes predict macrovascular outcomes indicating that the two types of outcome are related [18, 41] with possible common underlying pathophysiology [42, 43]. Dividing the major outcomes into micro-and macrovascular specific outcomes for prognostic models was demonstrated recently in type 2 diabetes patients [44]. The prognostic factors and risk estimates varied somewhat across type of complication. Unfortunately, the much younger age of type 1 diabetes patients with less events during the first 7 years of follow-up did not allow examination of major outcomes separately.

Further, participants in most of the studies were included between 1988 and 2002. Only the FinnDiane cohort had a very recent inclusion period (enrolment up to 2009). Even though evaluation in the FinnDiane cohort confirmed the validity of the risk estimates to more recent times, we cannot exclude that the frequency of major outcomes is different nowadays. Furthermore, we could not examine the generalizability of the prognostic model to non-white populations, since such cohorts were not available.

In conclusion, a prognostic model is now available to assess the absolute risk of major outcomes in patients with type 1 diabetes. The prognostic model may be useful for providing individual risk estimates of major outcomes. The risk estimates can guide surveillance recommendations, inform patients and allow efficient design and analysis of clinical trials.

Abbreviations

AER

albumin excretion rate

CACTI

Coronary Artery Calcification in Type 1 Diabetes study

CHD

coronary heart disease

C-statistic

concordance statistic

EDC

Epidemiology of Diabetes Complications Study

EURODIAB PCS

EURODIAB Prospective Complications Study

FinnDiane

Finnish Diabetic Nephropathy Study

HDL

high-density lipoprotein

WHR

waist-hip ratio