Comparison between data-driven clusters and models based on clinical features to predict outcomes in type 2 diabetes: nationwide observational study

Abstract

Aims/hypothesis Research using data-driven cluster analysis has proposed five novel subgroups of diabetes based on six measured variables in individuals with newly diagnosed diabetes. Our aim was (1) to validate the existence of differing clusters within type 2 diabetes, and (2) to compare the cluster method with an alternative strategy based on traditional methods to predict diabetes outcomes. Methods We used data from the Swedish National Diabetes Register and included 114,231 individuals with newly diagnosed type 2 diabetes. k-means clustering was used to identify clusters based on nine continuous variables (age at diagnosis, HbA1c, BMI, systolic and diastolic BP, LDL- and HDL-cholesterol, triacylglycerol and eGFR). The elbow method was used to determine the optimal number of clusters and Cox regression models were used to evaluate mortality risk and risk of CVD events. The prediction models were compared using concordance statistics. Results The elbow plot, with values of k ranging from 1 to 10, showed a smooth curve without any clear cut-off points, making the optimal value of k unclear. The appearance of the plot was very similar to the elbow plot made from a simulated dataset consisting only of one cluster. In prediction models for mortality, concordance was 0.63 (95% CI 0.63, 0.64) for two clusters, 0.66 (95% CI 0.65, 0.66) for four clusters, 0.77 (95% CI 0.76, 0.77) for the ordinary Cox model and 0.78 (95% CI 0.77, 0.78) for the Cox model with smoothing splines. In prediction models for CVD events, the concordance was 0.64 (95% CI 0.63, 0.65) for two clusters, 0.66 (95% CI 0.65, 0.67) for four clusters, 0.77 (95% CI 0.77, 0.78) for the ordinary Cox model and 0.78 (95% CI 0.77, 0.78) for the Cox model with splines for all variables. Conclusions/interpretation This nationwide observational study found no evidence supporting the existence of a specific number of distinct clusters within type 2 diabetes. The results from this study suggest that a prediction model approach using simple clinical features to predict risk of diabetes complications would be more useful than a cluster sub-stratification. Graphical abstract

Assessing reproducibility and utility of clustering of patients with type 2 diabetes and established CV disease (SAVOR -TIMI 53 trial)

“…Thus, to identify the four subtypes defined by Ahlqvist et al, we have chosen k = 4 when conducting k-means clustering. This approach is consistent with others performed similar analyses with type 2 diabetes cohorts [3,5,8,12].…”

A novel diabetes typology: towards precision diabetology from pathogenesis to treatment

“…An alternative strategy in precision diabetology may be based on statistical models using continuous risk factors [28,58,68]. In a secondary analysis of RCTs [28], age at diabetes diagnosis and renal function at baseline were better predictors of disease progression than the subgroup assignment according to Ahlqvist et al [6].…”

Type 2 diabetes classification: a data-driven cluster study of the Danish Centre for Strategic Research in Type 2 Diabetes (DD2) cohort

“…Similarly, other studies did not find their MOD-like clusters to have the highest BMI 15 22 or did not support k=4 as the most optimal number of T2D clusters. 16 …”

“…Other studies also have striven to identify novel T2D clusters in Asian,14 15 18 19 Latin American,19 and Caucasian16 19–21 diabetes patients using k-means cluster analysis. Some have used a fixed k-value of four in order to replicate the original Swedish clusters,17 18 whereas others14–16 20 21 have applied an analytical approach allowing for a wider range of clusters and/or included cluster variables other than those originally used. Our DD2 MOD cluster seemed less well defined, as patients from this cluster reallocated to all three new alternative k-value DD2 clusters.…”

“…To allow data to determine the number of clusters, rather than forcing a k-value of 4, we visually evaluated the optimal number of clusters. We selected the average silhouette method a priori because it was used in most previous diabetes cluster analyses2 14–16 and is considered the most objective method. We then reran the de novo k-means cluster analysis on the DD2 cohort using the alternative k-value.…”