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

Area of Science:

• Endocrinology
• Metabolic Diseases
• Biostatistics

Background:

• Previous research suggested novel subgroups of diabetes using cluster analysis.
• Type 2 diabetes (T2D) is a complex metabolic disorder with varying clinical presentations and outcomes.

Purpose of the Study:

• To validate distinct clusters within type 2 diabetes using data-driven cluster analysis.
• To compare cluster-based prediction with traditional methods for diabetes outcomes.

Main Methods:

• Utilized data from the Swedish National Diabetes Register (114,231 individuals with newly diagnosed T2D).
• Applied k-means clustering based on nine continuous variables (e.g., age, HbA1c, BMI, BP, lipids, eGFR).
• Employed Cox regression models to assess mortality and cardiovascular disease (CVD) event risks, comparing cluster models with standard Cox models.

Main Results:

• The elbow method did not clearly identify an optimal number of clusters, showing a smooth curve similar to a single-cluster dataset.
• Cluster-based models showed lower predictive accuracy (concordance 0.63-0.66) compared to ordinary Cox models (concordance 0.77) and spline-adjusted Cox models (concordance 0.78) for mortality and CVD events.

Conclusions:

• No evidence supports distinct, quantifiable clusters within type 2 diabetes in this large cohort.
• Predictive models utilizing standard clinical features offer superior utility for risk stratification of diabetes complications over cluster sub-stratification.