Ismail Saudi¹*, Abqariyah Yahya1, Sanjay Rampal1
1. Department of Social and Preventive Medicine, Faculty of Medicine, Universiti Malaya, 50603, Kuala Lumpur, Malaysia
*Corresponding author: Dr Ismail Bin Saudi.
ABSTRACT
INTRODUCTION: The clinical heterogeneity of Type 2 Diabetes Mellitus (T2DM) challenges standardised management. Data-driven subtyping using real-world data is crucial for advancing precision medicine, yet such analyses are absent for the diverse Malaysian population, representing a critical knowledge gap.
OBJECTIVE(S): To identify and characterise distinct clinical phenotypes of T2DM using an unsupervised machine learning approach on data from the Malaysian National Diabetes Registry.
MATERIALS & METHODS: This cross-sectional study analysed 123,137 adults with T2DM. Six routinely collected variables were used: age at diagnosis, diabetes duration, body mass index, HbA1c, triglyceride-glucose index, and triglyceride-to-HDL ratio. Following rigorous preprocessing including clinically informed outlier handling and random forest-based imputation, data were standardised and partitioned using K-means clustering. A comprehensive validation framework, incorporating statistical metrics, bootstrap analysis for stability, and confirmation with hierarchical clustering and Gaussian mixture models, was employed to ensure the robustness of the final cluster solution.
RESULTS: Four clinically distinct and robust T2DM phenotypes were identified. These were: Severe Insulin-Resistant Diabetes (SIRD, 24%) with high insulin resistance and long disease duration; Severe Insulin-Deficient Diabetes (SIDD, 18%) with the poorest glycaemic control (mean HbA1c 10.4%); Mild Age-Related Diabetes (MARD, 29%) characterised by late onset and the best metabolic profile; and a large Obesity-Related Diabetes group (ORD, 30%) with the highest BMI and severe insulin resistance. All patient characteristics differed significantly across clusters (p<0.001).
CONCLUSION: This first national-level cluster analysis reveals four distinct T2DM phenotypes within the Malaysian population. The identification of these subgroups, particularly a severe obesity-related phenotype, provides a foundational evidence base for developing targeted screening, prevention, and treatment strategies. This work represents a critical first step towards implementing precision diabetology in Malaysia.
KEYWORDS: Diabetes Clustering, Machine Learning, Malaysia, Precision Medicine, Registry Data


