Validation of a treatment selection algorithm for optimal choice of SGLT2 and DPP4 inhibitor therapies in people with type 2 diabetes across major UK ethnicity groups

Background Targeting type 2 diabetes treatments based on the routine clinical features of individual patients represents a practical precision medicine approach with potential worldwide applicability. We aimed to evaluate a recently proposed treatment selection model predicting differences in glycaemic responses to SGLT2-inhibitors and DPP4-inhibitors across major UK ethnicity groups. Methods We externally validated the SGLT2i-DPP4i model in UK primary care cohort (CPRD Aurum, 2013-2020) independent of the original model development cohort. Non-insulin treated individuals with type 2 diabetes were identified and categorised by major UK ethnicity groups: White, Black, South Asian and Mixed/Other. For each ethnicity group, we applied a closed testing procedure to assess whether model recalibration was required. After model updates, we assessed the calibration accuracy of predicted differences in glycaemic response (6-month change in HbA1c) between SGLT2i and DPP4i for each ethnicity group. Results SGLT2i (n=57,749) and DPP4i (n=87,807) initiations were identified amongst people of White (n=114,287; 78.5%), Black (n=6,663; 4.6%), South Asian (n=20,969; 14.4%) and Mixed/Other (n=3,637; 2.5%) ethnicities. Minor model adjustment was required to adjust for greater observed than predicted glycaemic responses to DPP4i (White -1.6 mmol/mol; Black -3.0 mmol/mol; South Asian -2.6 mmol/mol; Mixed/Other -2.6 mmol/mol). SGLT2i predictions did not require adjustment for non-White ethnicity groups. After model updates, average predicted HbA1c reduction was 3.7 mmol/mol (95%CI 3.5-3.9) greater with SGLT2i than DPP4i for those of White ethnicity; this was greater than for those of South Asian (2.1 mmol/mol (95%CI 1.6-2.6)), Black (0.6 mmol/mol (95%CI 0.5-1.7)) and Mixed/Other (2.6 mmol/mol (95%CI 1.4-3.8)) ethnicity groups. For all ethnicity groups, calibration of predicted differential glycaemic treatment effects were well-calibrated. Conclusion A treatment selection model for SGLT2-inhibitor and DPP4-inhibitor therapies is accurate for all major UK ethnicity groups. Simple recalibration is beneficial to optimise performance and this is recommended prior to deployment of the model in new populations and settings.