A five-drug class model using routinely available clinical features to optimise prescribing in type 2 diabetes: a...

Data to support individualised choice of optimal glucose-lowering therapy are scarce for people with type 2 diabetes. We aimed to establish whether routinely available clinical features can be used to predict the relative glycaemic effectiveness of five glucose-lowering drug classes.

We developed and validated a five-drug class model to predict the relative glycaemic effectiveness, in terms of absolute 12-month glycated haemoglobin (HbA_1c), for initiating dipeptidyl peptidase-4 inhibitors, glucagon-like peptide-1 receptor agonists, sodium-glucose co-transporter-2 inhibitors, sulfonylureas, and thiazolidinediones. The model used nine routinely available clinical features of people with type 2 diabetes at drug initiation as predictive factors (age, duration of diabetes, sex, and baseline HbA_1c, BMI, estimated glomerular filtration rate, HDL cholesterol, total cholesterol, and alanine aminotransferase). The model was developed and validated with observational data from England (Clinical Practice Research Datalink [CPRD] Aurum), in people with type 2 diabetes aged 18-79 years initiating one of the five drug classes between Jan 1, 2004, and Oct 14, 2020, with holdback validation according to geographical region and calendar period. The model was further validated in individual-level data from three published randomised drug trials in type 2 diabetes (TriMaster three-drug crossover trial and two parallel-arm trials [NCT00622284 and NCT01167881]). For validation in CPRD, we assessed differences in observed glycaemic effectiveness between matched (1:1) concordant and discordant groups receiving therapy that was either concordant or discordant with model-predicted optimal therapy, with optimal therapy defined as the drug class with the highest predicted glycaemic effectiveness (ie, lowest predicted 12-month HbA_1c). Further validation involved pairwise drug class comparisons in all datasets. We also evaluated associations with long-term outcomes in model-concordant and model-discordant groups in CPRD, assessing 5-year risks of glycaemic failure (confirmed HbA_1c ≥69 mmol/mol), all-cause mortality, major adverse cardiovascular events or heart failure (MACE-HF) outcomes, renal progression, and microvascular complications using Cox proportional hazards regression adjusting for relevant demographic and clinical covariates.

The five-drug class model was developed from 100 107 drug initiations in CPRD. In the overall CPRD cohort (combined development and validation cohorts), 32 305 (15·2%) of 212 166 drug initiations were of the model-predicted optimal therapy. In model-concordant groups, mean observed 12-month HbA_1c benefit was 5·3 mmol/mol (95% CI 4·9-5·7) in the CPRD geographical validation cohort (n=24 746 drug initiations, n=12 373 matched pairs) and 5·0 mmol/mol (4·3-5·6) in the CPRD temporal validation cohort (n=9682 drug initiations, n=4841 matched pairs) compared with matched model-discordant groups. Predicted HbA_1c differences were well calibrated with observed HbA_1c differences in the three clinical trials in pairwise drug class comparisons, and in pairwise comparisons of the five drug classes in CPRD. 5-year risk of glycaemic failure was lower in model-concordant versus model-discordant groups in CPRD (adjusted hazard ratio [aHR] 0·62 [95% CI 0·59-0·64]). For long-term non-glycaemic outcomes, model-concordant versus model-discordant groups had a similar 5-year risk of all-cause mortality (aHR 0·95 [0·83-1·09]) and lower risks of MACE-HF outcomes (aHR 0·85 [0·76-0·95]), renal progression (aHR 0·71 [0·64-0·79]), and microvascular complications (aHR 0·86 [0·78-0·96]).

We have developed a five-drug class model that uses routine clinical data to identify optimal glucose-lowering therapies for people with type 2 diabetes. Individuals on model-predicted optimal therapy had lower 12-month HbA_1c, were less likely to need additional glucose-lowering therapy, and had a lower risk of diabetes complications than individuals on non-optimal therapy. With setting-specific optimisation, the use of routinely collected parameters means that the model is easy to introduce to clinical care in most countries worldwide.

UK Medical Research Council.