Identification and epidemiological characterization of Type-2 diabetes sub-population using an unsupervised machine learning approach

Background: Type-2 Diabetes Mellitus (T2DM) exhibits heterogeneous sub-populations, but identification of sub-populations within epidemiological datasets is underexplored. This study focuses on detecting T2DM clusters in the Indian National Family Health Survey-4 (NFHS-4) dataset containing 10,125 T2DM patients with diverse features spanning medical history, diet, addictions, socio-economic, and lifestyle factors. Methods: Conventional UMAP for dimensionality reduction performed poorly due to mixed data types. The authors implemented a feature-type-distributed workflow applying UMAP separately to continuous, ordinal, and nominal features using appropriate similarity metrics (Euclidean for continuous, Canberra for ordinal, Hamming for nominal). The resulting low-dimensional embeddings (2D for continuous and ordinal; 1D for nominal) were integrated into a five-dimensional representation and clustered using DBSCAN. Results: Seven clusters (with 261 outliers) were detected; four significant clusters contained 2898, 2301, 2226, and 1315 individuals. Two clusters were predominantly non-obese, characterized by lower mean age and BMI and a higher proportion of rural residents and lower wealth status. One cluster showed very low non-vegetarian intake (about 90% reporting no egg, fish, chicken/meat intake). Conclusions: Feature-type-distributed UMAP clustering is effective for heterogeneous epidemiological data. The presence of significant non-obese T2DM sub-populations with younger age and economic disadvantage highlights the need for different T2DM screening criteria among rural inhabitants and tailored dietary guidance.