○ Type 1 diabetes. Best diagnostic results depend on integrating all diagnostic modalities, not by relying on prior prevalence, clinical features or test results in isolation. The age at which the initial islet autoantibody appears and the type of autoantibody (e.g. which of the four primary antibodies among ICA512, insulin, GAD and ZnT8) may be important in defining aetiological subtypes of type 1 diabetes. The majority of the genetic risk of type 1 diabetes is now known, and the sensitivity and specificity of a type 1 diabetes genetic risk score (T1D-GRS) both exceed 80%. Despite this, a high T1D-GRS will have low positive predictive value in patient populations where the overall prevalence of type 1 diabetes is low, such as those aged >50 years when diabetes is diagnosed. It will likely prove most useful when the T1D-GRS is combined with clinical features and islet autoantibodies. At present, there is no immune-based test sufficiently reproducible and robust that it can be used diagnostically

○ Type 2 diabetes. Cluster analysis at diagnosis can provide insights into likely progression, risk of complications, and treatment response, which offer an exciting approach to subclassification of type 2 diabetes. At this time, the available genetic data for type 2 diabetes do not have sufficient predictive accuracy to replace existing delineative approaches. Although the subcategorisation of type 2 diabetes using genetic data are informative regarding the aetiological processes that underlie the disease, the methods described so far [6, 102] are not intended to be used to subclassify a type 2 diabetes diagnosis nor are the existing genetic data sufficient for this purpose for the majority of individuals with type 2 diabetes. Treatment response and progression can be predicted from clinical features [138]. An advantage of using clinical features for diagnosis of type 2 diabetes is that they are widely available and easily obtained (e.g. sex, BMI, HbA1c); however, a potential limitation is that they may vary over time.

○ Barriers to implementation. One of several important translational barriers facing the proposed clustering approach for type 1 and type 2 diabetes is that a fasting C-peptide measurement is required at the time of diagnosis, which is not routinely performed in clinical practice, and the reliability of C-peptide assays vary considerably between laboratories [42. Another limitation is that the biomarkers used to define these clusters change over time depending on the disease course or its treatment, such that this approach can only be applied to newly diagnosed individuals, but not to individuals years before disease onset or the many millions of people with long-standing diabetes worldwide. Moreover, because the current approaches for clustering in type 2 diabetes require continuously distributed data to be categorised, which typically results in loss of power. Thus, these methods do not yield good predictive accuracy, a major expectation in precision medicine, but this may change as the approach is refined.

○ Research gaps. Based on limited ideal tests and uncertainty in aetiology, more research is needed on type 1 and type 2 diabetes in order to define subtypes and decide the best interventional and therapeutic approaches.