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

Summary

Using UK CPRD data, the authors developed and validated a model that predicts the most effective of five glucose-lowering drug classes for individual patients based on routine clinical features. Only 15.2% of initiations aligned with model-optimal therapy; patients receiving model-concordant therapy achieved lower 12‑month HbA1c than those on discordant therapy.

Key Findings

• A five-drug class prediction model was developed and validated using 100,107 drug initiations in CPRD.
• Across 212,166 initiations, only 15.2% matched the model-predicted optimal therapy.
• Model-concordant treatment was associated with lower observed 12-month HbA1c than discordant treatment.

Clinical Implications

Embedding the model into EHR decision support could guide initial or add-on therapy selection (e.g., between SGLT2 inhibitors, GLP-1 receptor agonists, DPP-4 inhibitors, TZDs, sulfonylureas) to maximize HbA1c reduction, potentially reducing therapeutic inertia and unnecessary switching.

Limitations

• Observational design may entail residual confounding and indication bias
• Generalizability beyond UK settings and non-glycemic outcomes (e.g., CV-renal endpoints) require further study

Future Directions

Prospective pragmatic trials testing clinical outcomes, health-economic evaluations, and deployment as CDS tools with continuous recalibration across diverse populations.