Type 1 Diabetes Risk Phenotypes Using Cluster Analysis

Abstract

Background: Although statistical models for predicting type 1 diabetes risk have been developed, approaches that reveal clinically meaningful clusters in the at-risk population and allow for non-linear relationships between predictors are lacking. We aimed to identify and characterize clusters of islet autoantibody-positive individuals that share similar characteristics and type 1 diabetes risk.

Methods: We tested a novel outcome-guided clustering method in initially non-diabetic autoantibody-positive relatives of individuals with type 1 diabetes, using the TrialNet Pathway to Prevention (PTP) study data (n=1127). The outcome of the analysis was time to type 1 diabetes and variables in the model included demographics, genetics, metabolic factors and islet autoantibodies. An independent dataset (Diabetes Prevention Trial of Type 1 Diabetes, DPT-1 study) (n=704) was used for validation.

Findings: The analysis revealed 8 clusters with varying type 1 diabetes risks, categorized into three groups. Group A had three clusters with high glucose levels and high risk. Group B included four clusters with elevated autoantibody titers. Group C had three lower-risk clusters with lower autoantibody titers and glucose levels. Within the groups, the clusters exhibit variations in characteristics such as glucose levels, C-peptide levels, age, and genetic risk. A decision rule for assigning individuals to clusters was developed. The validation dataset confirms that the clusters can identify individuals with similar characteristics.

Interpretation: Demographic, metabolic, immunological, and genetic markers can be used to identify clusters of distinctive characteristics and different risks of progression to type 1 diabetes among autoantibody-positive individuals with a family history of type 1 diabetes. The results also revealed the heterogeneity in the population and complex interactions between variables.

Figure 1.

Dendrogram and heatmap resulted from the clustering analysis. From left to right, the displayed information is the complete-linkage dendrogram (Column 1), the heatmap of demographic information (Column 2), the heatmap of genetic information (Column 3), the heatmap of immunological markers (Column 4), the heatmap of metabolic markers (Column 5), the heatmap of T1D outcomes and rates (Column 6), diabetes-free Kaplan-Meier survival curves of clusters (Column 7), distribution of clusters on the multidimensional scaling map (Column 8), and, in the foremost right side, the legends denoting the color scheme for the variables (Column 9).

Figure 2.

MDS map of clusters (center), flanked by the radar charts of each cluster (surrounding the MDS maps), and the Kaplan-Meier survival curves by cluster with their corresponding 95% confidence intervals (outer).

Figure 3.

Decision rules to assign individuals to clusters.

Figure 4.

Comparison of diabetes-free Kaplan-Meier survival curves of clusters in the analysis dataset (TrialNet PTP; black solid lines) and the validation dataset (DPT-1; blue dashed lines). The numbers under the curves are the numbers of subjects at risk in the cluster at each time point P-values on the upper right corners are based on log-rank tests.