A Classification Algorithm for Diabetes Subtypes in a Swedish Cohort: Comparison With National Diabetes Register Diagnoses and Genetic Validation

To address limitations in diabetes subtype diagnoses recorded in registers, we aimed to systematically classify 18,985 individuals from the All New Diabetics in Scania (ANDIS) cohort and compare them with Swedish National Diabetes Register (NDR) diagnoses. Type 1 and type 2 diabetes polygenic risk scores (PRS1 and PRS2, respectively) were used for genetic validation. We developed a classification algorithm using linked longitudinal data from regional health care databases, national registers, and cohort-based assessments, guided by American Diabetes Association/European Association for the Study of Diabetes guidelines. In ANDIS, we classified 3.3% of subtypes as type 1 diabetes, 92% as type 2 diabetes, 3.6% as latent autoimmune diabetes in adults (LADA), 0.5% as pancreatogenic (type 3c) diabetes, and 0.5% as unspecified diabetes. Among 16,591 individuals with NDR-reported diagnoses, 3.9% received a different diabetes subtype, showing some residual disagreement between diagnoses despite a substantial agreement overall (κ = 0.71). PRSs independently validated our classification, separating type 1 diabetes, type 2 diabetes, and LADA and demonstrating a consistent genetic gradient among reclassified individuals compared with NDR. Clinical-genetic analyses suggested that most unspecified diabetes did not resemble type 1 diabetes, LADA, or type 3c diabetes. Our algorithm reliably distinguishes diabetes subtypes with higher granularity compared with NDR, reveals heterogeneity overlooked in registers, and enables subtype-specific research. Article Highlights Diabetes subtype–specific research is challenging when using data from large registers, and validation studies of Swedish National Diabetes Register (NDR) diagnoses remain limited. We aimed to develop an algorithm to classify diabetes subtypes using linked longitudinal data, compare with NDR, and validate it using polygenic risk scores. We developed an algorithm that reliably distinguishes diabetes subtypes with higher granularity compared with NDR and is validated by polygenic risk scores. Our algorithm reveals clinically meaningful heterogeneity that may be overlooked in registers and enables subtype-specific diabetes research.