Abstract
AIMS/HYPOTHESIS: Despite continued interest in precision diagnostics and type 2 diabetes subtypes, the challenge of uncertainty in the classification of individuals into subtypes remains. This study introduces a novel method for quantifying and accounting for classification uncertainty in type 2 diabetes subtypes. METHODS: Building on recommendations from the ADA/EASD Precision Medicine in Diabetes Initiative, we quantified classification uncertainty using the normalised relative entropy (NRE), computed from distances to cluster centroids. A lower NRE value indicates greater uncertainty in an individual's cluster assignment. We examined the NRE in a cohort of 859 individuals with recent-onset type 2 diabetes from the prospective, observational German Diabetes Study (GDS) and compared it across previously identified diabetes subtypes, defined by age, BMI, HbA(1c), HOMA-IR and HOMA-B. Predicted 10 year CVD risk (SCORE2-Diabetes) of the subtypes was evaluated with and without accounting for classification uncertainty. RESULTS: Individuals with mild age-related diabetes (n=395) and mild obesity-related diabetes (n=316) had a median NRE of 0.155 (95% CI 0.142, 0.177) and 0.119 (95% CI 0.107, 0.131), respectively. By contrast, individuals with severe insulin-resistant diabetes (n=130) and severe insulin-deficient diabetes (n=18) had a lower median NRE of 0.086 (95% CI 0.075, 0.108) and 0.082 (95% CI 0.071, 0.109), respectively. After weighting individuals by classification certainty, the proportion of variation in SCORE2-Diabetes explained by the subtypes (R(2)) increased from 17.4% (95% CI 12.8, 23.0) to 31.5% (95% CI 26.4, 37.1). The predicted 10 year CVD risk of the mild age-related diabetes subtype increased from 10.3% (95% CI 9.8, 10.7) to 11.6% (95% CI 11.2, 12.0). CONCLUSIONS/INTERPRETATION: The NRE provides a means to quantify and compare individual classification uncertainty in type 2 diabetes subtypes. Classification uncertainty varied between subtypes and individuals with type 2 diabetes, and accounting for it improved the ability of the subtypes to predict 10 year CVD risk.