CSEMO1 - Machine Learning-Derived Diabetes Clusters and Risk of Diabetes Complications: A Systematic Review and Meta-analysis

Abstract:

Background: In 2018, a seminal study by Ahlqvist et al. identified five distinct diabetes clusters using machine learning (ML). Since then, multiple groups have validated these clusters and reported different rates of diabetes-related outcomes in each cluster. This systematic review aims to describe all studies reporting diabetes-related outcomes in the Ahlqvist clusters.

Methods: We conducted a systematic review of OVID Medline, EMBASE, and Cochrane Library from inception to June 2025. Studies using ML algorithms to classify patients with diabetes into Ahlqvist clusters and reporting on diabetes-related clinical outcomes were included. Population demographics, diabetes-related variables, and clinical outcomes were recorded.

Results: Of 1,386 studies, 26 met inclusion criteria, comprising 54,307 patients. Mean age was 61.4 years and 46.8% were female. De novo k-means clustering was done by 17 studies (65.4%), followed by a nearest centroid approach based on an Ahlqvist cohort (19.2%). Clustering revealed 7.4% with severe autoimmune diabetes (SAID), 19.9% severe insulin deficient diabetes (SIDD), 16.5% with severe insulin resistant diabetes (SIRD), 27.2% with mild obesity-related diabetes (MOD), and 32.8% with mild age-related diabetes (MARD). The SIRD cluster was most strongly associated with nephropathy and macrovascular complications (20 and 9 studies; 76.9% and 34.6%, respectively) while SIDD patients were at highest risk for retinopathy (11 studies, 42.3%).

Conclusion: The Ahlqvist clustering framework has been widely replicated in the literature. Our review revealed associations between specific clusters with known complications of diabetes. These findings highlight the potential of ML-based clustering to identify new subgroups of diabetes and personalize diabetes management.