Predicting renal disease progression in a large contemporary cohort with type 1 diabetes mellitus

Aims/hypothesis The aim of this study was to provide data from a contemporary population-representative cohort on rates and predictors of renal decline in type 1 diabetes. Methods We used data from a cohort of 5777 people with type 1 diabetes aged 16 and older, diagnosed before the age of 50, and representative of the adult population with type 1 diabetes in Scotland (Scottish Diabetes Research Network Type 1 Bioresource; SDRNT1BIO). We measured serum creatinine and urinary albumin/creatinine ratio (ACR) at recruitment and linked the data to the national electronic healthcare records. Results Median age was 44.1 years and diabetes duration 20.9 years. The prevalence of CKD stages G1, G2, G3 and G4 and end-stage renal disease (ESRD) was 64.0%, 29.3%, 5.4%, 0.6%, 0.7%, respectively. Micro/macroalbuminuria prevalence was 8.6% and 3.0%, respectively. The incidence rate of ESRD was 2.5 (95% CI 1.9, 3.2) per 1000 person-years. The majority (59%) of those with chronic kidney disease stages G3–G5 did not have albuminuria on the day of recruitment or previously. Over 11.6 years of observation, the median annual decline in eGFR was modest at −1.3 ml min−1 [1.73 m]−2 year−1 (interquartile range [IQR]: −2.2, −0.4). However, 14% experienced a more significant loss of at least 3 ml min−1 [1.73 m]−2. These decliners had more cardiovascular disease (OR 1.9, p = 5 × 10−5) and retinopathy (OR 1.3 p = 0.02). Adding HbA1c, prior cardiovascular disease, recent mean eGFR and prior trajectory of eGFR to a model with age, sex, diabetes duration, current eGFR and ACR maximised the prediction of final eGFR (r2 increment from 0.698 to 0.745, p < 10−16). Attempting to model nonlinearity in eGFR decline or to detect latent classes of decliners did not improve prediction. Conclusions These data show much lower levels of kidney disease than historical estimates. However, early identification of those destined to experience significant decline in eGFR remains challenging. Electronic supplementary material The online version of this article (10.1007/s00125-019-05052-z) contains peer-reviewed but unedited supplementary material, which is available to authorised users.

Using 5 years data retrospective to study day (the day when participants were recruited into SDRNT1BIO), we estimated a summary trajectory of eGFR separately for each individual. To accomplish that, within each month we first considered the median eGFR readings non-concurrent with hospital admissions. To these data we applied an exponential smoothing with a time window of 1 year according to: where s i is the i-th smoothed eGFR reading for that individual, and w i = max(1 − d i /365, 0) is the weight assigned to the previous terms, which depends on the number of days d i between two successive eGFR readings, eGFR i−1 and eGFR i .
We then fitted a linear regression model for each participant over the smoothed data, as long as the individual had at least 3 readings over a period of at least 2 years. Therefore, for each person with n smoothed eGFR observations recorded at time t i (expressed in years from the first reading used), we fitted the following model: where α and β are intercept and slope, respectively, and ε i ∼ N (0, σ 2 ) is the random error. Thus, the slope term β is the average annualized effect of time on eGFR for that individual.
To assess the frequency of trajectories that are non-linear, we fitted for each person a model containing a quadratic term for time, of the form: To examine across all individuals the extent to which taking into consideration any departure from linearity is useful for prediction of final eGFR, we compared the adjusted r 2 with respect to the baseline model for a linear regression model of final eGFR with and without a quadratic term across all individuals.
We compared the above approach of computing slopes to using a linear mixed model (LMM) and conditional two-step LMM [2], as well as a linear mixed effects model with non-stationary stochastic processes [3], as implemented in the lmenssp R package (version 1.2: https://CRAN.R-project.org/package=lmenssp).
We further formulated the problem as a latent class mixed model, which allows to assign group membership to each participant according to their profile of trajectory. We attempted to fit models with 2 and 3 latent classes using the R package lcmm [4] (version 1.7.8: https://CRAN.R-project.org/package=lcmm).

ESM Results
To test whether eGFR decline can be treated as being linear we fitted individual models to test whether quadratic terms improve the fit of eGFR trajectories beyond simple linear terms. In 18.9% there was some improvement in fit of the model of trajectories as evidenced by the quadratic term being statistically significant at p < 0.01 (26.2% of participants with an average loss in eGFR of at least 3 ml min −1 [1.73m] −2 year −1 , and 15.7% of the remainder). However, when we evaluated how useful including a quadratic term is across all individuals for predicting final eGFR, the improvement in prediction was trivial (r 2 increased from 0.698 to 0.701 only). This was the case both in those with study day eGFR above or below 60 ml min −1 [1.73m] −2 , thus the remaining analyses did not include quadratic effects of time on eGFR.
Neither using linear mixed models with or without stochastic processes improved prediction performance. Using a latent class approach to establish the existence of a group with moderate or fast decline, no classes were discoverable.

ESM References
[ We report frequency (as %) for categorical variables and median (IQR) for continuous variables ESM