To Cluster or Not to Cluster? A Comparison of Approaches to Targeting Type 2 Diabetes Glucose-Lowering Therapy in the...

Different precision treatment approaches have been proposed for type 2 diabetes, but robust comparisons are lacking. We compared the utility of proposed approaches targeting glucose-lowering therapy in the TriMaster three-way crossover trial.

We evaluated four previously reported precision treatment strategies for their ability to predict overall 4-month HbA1c response and within-person differential response to sitagliptin, canagliflozin, and pioglitazone in 309 adults with type 2 diabetes who received all three medications. Approaches included allocation to clusters based on routine features and HOMA, direct outcome prediction using a routine-features treatment-selection model, type 2 diabetes cluster-specific partitioned polygenic scores (PPSs), and a model based on discriminative dimensionality reduction tree (DDRTree) analysis.

The routine-features model and cluster approach were both strongly associated with overall and differential HbA1c responses (P < 0.0001). The routine-features model identified the most effective therapy with significant benefit across all drug comparisons, whereas clinical clusters showed benefit for only one cluster-drug comparison (severe insulin-resistant diabetes: SGLT2i vs. thiazolidinedione [TZD], P = 0.005). DDRTree was not associated with treatment response. Of eight PPSs, only two showed modest association with differential response (lipodystrophy PPS had a greater response to SGLT2 vs. TZD; and β-cell dysfunction, negative proinsulin PPS had a greater response to TZD vs. dipeptidyl peptidase 4 inhibitor). Treatment allocation based on the routine-features model resulted in greater HbA1c reduction than cluster-based allocation (0.27% [95% CI 0.18, 0.35], 2.9 mmol/mol [95% CI 2.0, 3.8] vs. 0.16% [95% CI 0.07, 0.25], 1.7 mmol/mol [95% CI 0.8, 2.7]).

Clinical features are more strongly associated with differential glycemic response than PPS, and direct outcome prediction outperforms data-driven subgrouping for optimizing short-term HbA1c reduction.