Risk-Based Clustering for Patient Stratification in Diabetes

This study explores the use of unsupervised clustering to identify clinically meaningful subgroups within a population-level diabetes dataset. Using six clinical features— age, BMI, HbA1c, blood glucose, hypertension, and heart disease—we applied Principal Component Analysis (PCA) for dimensionality reduction and K-Means clustering with two, three and four clusters. Clusters were evaluated using silhouette scores and the Elbow method. Cluster profiles were visualized through PCA scatter plots, bar plots, and heatmaps. We further assessed cluster validity by comparing subgroup characteristics to diabetes diagnosis labels and using logistic regression post hoc to identify the most predictive variables. Clusters with elevated HbA1c and blood glucose levels aligned strongly with diagnosed diabetes cases and also highlighted potentially undiagnosed individuals. Secondary variables such as gender and smoking status showed mild cluster variation. Overall, this approach demonstrates how unsupervised learning can reveal latent risk patterns and support early detection strategies in diabetes care.