We developed and validated a clinical prediction rule that can be used to predict progression from normal AER to microalbuminuria in type 1 diabetes patients. First, we identified a series of predictors for the risk of progression to microalbuminuria within a period of 7 years. The following five characteristics predominantly contributed to prediction in a multivariable logistic regression model: HbA1c, AER, WHR, BMI and ever smoking. These predictors can be relatively easily obtained by physical examination and laboratory tests. External validation of the prediction rule in three independent patient samples showed good performance.
and baseline AER have consistently been shown to be important predictors of microalbuminuria [2
] and were also included in a previous model developed by Rossing and co-workers [5
]. Smoking was also included in both models. In contrast, BMI and WHR were included in the present model, but were not considered in the development of the Rossing model, which did, however, include retinopathy. We found with the EURODIAB data that the present model discriminated patients with microalbuminuria from those with normal AER better than the Rossing model, which included retinopathy (c
-statistics 0.69 and 0.65 respectively). A model that included the four predictors of the Rossing model (HbA1c
, AER, current smoking and any retinopathy) fitted with the EURODIAB PCS data showed a c
-statistic of 0.66. The additive predictive value for WHR in a model that already includes BMI indicates that central obesity contains different information than general obesity. This is confirmed by the relatively low correlation between WHR and BMI (Pearson correlation r
0.21). The differences in effect of central and general obesity have also been described for the risk of cardiovascular events [23
The predictors hyperglycaemia and central obesity are part of the insulin resistance syndrome [24
]. Insulin resistance is also a risk factor for the development of diabetic complications. We further studied the added value of insulin sensitivity with the amount of fasting insulin (per kg body weight). However, this factor did not have added value in our model (p
Continuous predictors are best included in a model as such and not categorised [20
]. In this way, all information is used for the prediction. Continuous variables do not necessarily have to be included as linear terms. We found that logarithmic transformation of HbA1c
predicts the risk of microalbuminuria better than the frequently used linear values. The nature of the relation between AER and risk of microalbuminuria was linear, although a logarithmic transformation of AER is frequently used [3
]. Apparently, a logarithmic transformation of AER is only necessary to estimate correct values of the mean and standard deviation, given the skewed distribution of AER.
We studied the generalisability of our prediction rule with respect to place (EDC), time (FinnDiane) and time and place (CACTI). The US cohort (EDC) with patients treated in the same time period as the patients from the development set (the 1990s) showed similar discriminative ability of the prediction rule (c
0.71 compared with 0.69 in the development set). C
-statistic values around 0.7 indicate reasonable discriminative ability for prognostic models. Prediction of an event later in time (here 7 years) is more difficult than prediction of an event shortly after the baseline measurement. Furthermore, the predicted risks were reasonably in agreement with the observed proportions of microalbuminuria (Table ).
It was particularly important to study the generalisability of the prediction rule in time, since treatment of type 1 diabetes patients with normal AER has changed. Nowadays, ACE inhibitors and statins are prescribed more frequently to lower blood pressure and lipid plasma concentrations respectively. The use of these medications has been shown to affect the transition from normo- to microalbuminuria in type 1 diabetes patients [25
], as was also apparent from our data. The incidence of microalbuminuria was only 7.7% in FinnDiane and 6.0% in CACTI compared with 13% in the earlier development sample. Indeed, predicted risks were too high for the patients from the FinnDiane and CACTI studies. However, the model was well able to discriminate between patients with and without microalbuminuria. Therefore, a simple recalibration step was sufficient to make the rule valid for recently treated patients.
The clinical implications of our study are that risk of microalbuminuria can be established at an early stage and patient management tailored to risk levels. We categorised the patients into four risk groups with scores 2 to 10, 11 to 15, 16 to 20, and 21 and higher. The highest two risk groups together contained 77% (110/143) of all microalbuminuric patients of the development set and 79%, 80% and 84% of all microalbuminuric patients of the validation sets. We would advocate offering these high-risk patients more frequent check-ups than once a year, perhaps together with the quarterly routine visit and with an overnight or 24 h urine collection to measure the albumin:creatinine ratio. Current surveillance protocols recommend follow-up of patients more frequently than once a year only after an abnormal surveillance result. Under this protocol, a considerable proportion of new microalbuminuria cases will be recognised later than necessary. The identification, by our prediction rule, of a group of patients with normal AER and at high risk of progressing to microalbuminuria might facilitate the early introduction of ACE inhibitor or angiotensin receptor blocker therapy, if trials currently under way suggest benefit. Glycaemic control could also be intensified and other risk factors for microalbuminuria, e.g. smoking, BMI and WHR could be more strictly controlled than is usual. It is currently unknown whether such intervention strategies based on our prediction rule would be of benefit; this should be the focus of future research.
Unfortunately, the model predicts for one occasion only, i.e. after 7 years of follow-up. The assessment of the outcome only after 7 years has two implications. First, we do not have information on time of onset of microalbuminuria. Time-to-event analysis, e.g. with Cox proportional hazards regression analysis, was therefore not possible. Second, we do not know how many patients remain microalbuminuric and how many regress. Several studies have shown regression of microalbuminuria in type 1 diabetes patients [2
]. Cumulative incidences of regression vary between 13% and 56% and are mainly induced by intensive therapy. Only few patients with microalbuminuria undergo spontaneous regression that is permanent (around 15%) [2
]. Consequently, our rule may only slightly overestimate a patient’s risk of progression to microalbuminuria.
In conclusion, we developed a prediction rule to estimate the risk of progressing to microalbuminuria in individual type 1 diabetes patients. The rule was developed in a European cohort and externally validated in two US cohorts and another European cohort. We believe that this prediction rule could be used to divide patients into different risk categories. Such risk categories could guide surveillance recommendations and ultimately improve the prevention of long-term chronic complications.