Identifying type 1 and 2 diabetes in population level data: assessing the accuracy of published approaches

Aims Population datasets are increasingly used to study type 1 or 2 diabetes, and inform clinical practice. However, correctly classifying diabetes type, when insulin treated, in population datasets is challenging. Many different approaches have been proposed, ranging from simple age or BMI cut offs, to complex algorithms, and the optimal approach is unclear. We aimed to compare the performance of approaches for classifying insulin treated diabetes for research studies, evaluated against two independent biological definitions of diabetes type. Method We compared accuracy of thirteen reported approaches for classifying insulin treated diabetes into type 1 and type 2 diabetes in two population cohorts with diabetes: UK Biobank (UKBB) n=26,399 and DARE n=1,296. Overall accuracy and predictive values for classifying type 1 and 2 diabetes were assessed using: 1) a type 1 diabetes genetic risk score and genetic stratification method (UKBB); 2) C-peptide measured at >3 years diabetes duration (DARE). Results Accuracy of approaches ranged from 71%-88% in UKBB and 68%-88% in DARE. All approaches were improved by combining with requirement for early insulin treatment (<1 year from diagnosis). When classifying all participants, combining early insulin requirement with a type 1 diabetes probability model incorporating continuous clinical features (diagnosis age and BMI only) consistently achieved high accuracy, (UKBB 87%, DARE 85%). Self-reported diabetes type alone had high accuracy (UKBB 87%, DARE 88%) but was available in just 15% of UKBB participants. For identifying type 1 diabetes with minimal misclassification, using models with high thresholds or young age at diagnosis (<20 years) had the highest performance. An online tool developed from all UKBB findings allows the optimum approach of those tested to be selected based on variable availability and the research aim. Conclusion Self-reported diagnosis and models combining continuous features with early insulin requirement are the most accurate methods of classifying insulin treated diabetes in research datasets without measured classification biomarkers.