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

This paper investigates the heterogeneous sub-populations within Type-2 Diabetes Mellitus (T2DM) using the National Family Health Survey-4 (NFHS-4) dataset from India. A novel feature-type-distributed clustering workflow, combining different similarity measure settings of UMAP for continuous, ordinal, and nominal features, was implemented to overcome the limitations of conventional UMAP application on diverse epidemiological data. The analysis identified four significant T2DM clusters, two of which comprised mainly non-obese patients, characterized by younger age, rural residence, and lower socioeconomic status. One cluster showed a surprisingly high percentage of non-vegetarian dieters despite a lack of increased plant-based protein intake. These findings highlight the need for modified T2DM screening criteria in rural areas and underscore the importance of considering diverse data types in epidemiological studies.

Saptrash Bej, Jit Sarkar, Saikat Biswas, Pabitra Mitra, Partha Chakrabarti, Olaf Wolkenhauer