The efficacy of canagliflozin in diabetes subgroups stratified by data-driven clustering or a supervised machine learning method: a post hoc analysis of canagliflozin clinical trial data

Abstract

Aims/hypothesis: Data-driven diabetes subgroups have shown distinct clinical characteristics and disease progression, although there is a lack of evidence that this information can guide clinical decisions. We aimed to investigate whether diabetes subgroups, identified by data-driven clustering or supervised machine learning methods, respond differently to canagliflozin.

Methods: We pooled data from five randomised, double-blinded clinical trials of canagliflozin at an individual level. We applied the coordinates from the All New Diabetics in Scania (ANDIS) study to form four subgroups: mild age-related diabetes (MARD); severe insulin-deficient diabetes (SIDD); mild obesity-related diabetes (MOD) and severe insulin-resistant diabetes (SIRD). Machine learning models for HbA_1c lowering (ML-A1C) and albuminuria progression (ML-ACR) were developed. The primary efficacy endpoint was reduction in HbA_1c at 52 weeks. Concordance of a model was defined as the difference between predicted HbA_1c and actual HbA_1c decline less than 3.28 mmol/mol (0.3%).

Results: The decline in HbA_1c resulting from treatment was different among the four diabetes clusters (p_interaction=0.004). In MOD, canagliflozin showed a robust glucose-lowering effect at week 52, compared with other drugs, with least-squares mean of HbA_1c decline [95% CI] being 6.6 mmol/mol (4.1, 9.2) (0.61% [0.38, 0.84]) for sitagliptin, 7.1 mmol/mol (4.7, 9.5) (0.65% [0.43, 0.87]) for glimepiride, and 9.8 mmol/mol (9.0, 10.5) (0.90% [0.83, 0.96]) for canagliflozin. This superiority persisted until 104 weeks. The proportion of individuals who achieved HbA_1c <53 mmol/mol (<7.0%) was highest in sitagliptin-treated individuals with MARD but was similar among drugs in individuals with MOD. The ML-A1C model and the cluster algorithm showed a similar concordance rate in predicting HbA_1c lowering (31.5% vs 31.4%, p=0.996). Individuals were divided into high-risk and low-risk groups using ML-ACR model according to their predicted progression risk for albuminuria. The effect of canagliflozin vs placebo on albuminuria progression differed significantly between the high-risk (HR 0.67 [95% CI 0.57, 0.80]) and low-risk groups (HR 0.91 [0.75, 1.11]) (p_interaction=0.016).

Conclusions/interpretation: Data-driven clusters of individuals with diabetes showed different responses to canagliflozin in glucose lowering but not renal outcome prevention. Canagliflozin reduced the risk of albumin progression in high-risk individuals identified by supervised machine learning. Further studies with larger sample sizes for external replication and subtype-specific clinical trials are necessary to determine the clinical utility of these stratification strategies in sodium-glucose cotransporter 2 inhibitor treatment.

Data availability: The application for the clinical trial data source is available on the YODA website ( http://yoda.yale.edu/ ).

Keywords: Data-driven clusters; Diabetes; Machine learning; SGLT2i.