Randomised clinical trials provide unbiased databases for comparative effectiveness analyses to see which patients respond best to available interventions.

We evaluated patient-level data pooled from seven Phase 3 clinical trials with liraglutide, to examine responder subgroups as defined by those achieving a composite endpoint of A1C <7%, no weight gain and no hypoglycaemia (episodes requiring assistance or self-treated with PG <56 mg/dL [<3.1 mmol/L]) over 26 weeks.

Overall 34% of individuals on liraglutide 1.8 mg achieved the prespecified composite endpoint. Candidate predictor variables included baseline age, sex, ethnicity, BMI, A1C, beta-cell function, FPG, insulin resistance, previous treatments, and diabetes duration. Using recursive partitioning to create classification trees, baseline A1C was the most significant predictor, with a 46% probability of achieving the composite outcome when baseline A1C <8.5%, as opposed to 19% if baseline A1C ≥8.5% (p<0.0001). Subsequent splits (with p-values <0.05) produced a subgroup within patients with a baseline A1C <8.5% that was identified by previous treatment with diet or monotherapy, female sex, and diabetes duration <4.9 years increasing probability of success to 74%. Six homogeneous subgroups were identified with different probabilities of achieving the composite outcome (Fig).

In summary, recursive partitioning identified individual characteristics and subgroups of patients predicting the response to therapy. Such analyses may guide clinicians in individualising treatment approaches.