Type 2 diabetes is increasingly recognized as a heterogeneous syndrome rather than a single disease. Clinical phenotype-based clustering, exemplified by the Ahlqvist model, has demonstrated that subgroups defined by simple variables exhibit distinct complication risks and treatment responses, yet this approach is limited by information loss due to categorization, static classification, and poor transferability across ethnicities. Genome-wide association studies have identified more than 1,200 independent risk signals and revealed genetically driven pathophysiological clusters enriched in cell-type-specific regulatory regions. In addition, partitioned polygenic risk scores show significant associations with vascular complications across ancestries. Multi-omics integration—spanning metabolomics, transcriptomics, epigenomics, and the gut microbiome—has further uncovered subtype-specific signatures that single-omics approaches cannot capture, though the translational gap between complex omics panels and clinically actionable biomarkers remains substantial. Artificial intelligence (AI) and machine learning methods, including ensemble models, deep survival analysis, foundation models trained on continuous glucose monitoring data, and multimodal fusion architectures, are enabling the integration of high-dimensional, heterogeneous data for complication prediction with performance surpassing conventional risk calculators. Recent studies combining polygenic risk scores with clinical variables in deep-learning frameworks have demonstrated significant improvements in predicting cardiovascular and renal complications, with evidence that genetic risk modifies the benefit of standard interventions. However, most models have been developed in European-ancestry populations, and their predictive accuracy diminishes substantially when applied to East Asian populations, where nonobese phenotypes and beta-cell dysfunction predominate. Korea possesses unique strengths to address these challenges, including nationwide health-insurance data, population-based genomic cohorts, and regulatory-approved AI-based retinal biomarkers that could serve as platforms for integrating genomic information. Realizing AI-driven precision diabetes care will require the concurrent development of population-specific prediction models, prospective multi-institutional validation, and robust data-governance frameworks.
Jae-Seung Yun (Mon,) studied this question.