Heterogeneity in the treatment effect of sodium-glucose cotransporter-2 inhibitor on chronic kidney disease risk reduction: a prevalent new user cohort study applying double/debiased machine learning.

Abstract

Background: Despite strong evidence supporting the benefits of sodium-glucose cotransporter-2 inhibitor (SGLT2i) for managing type 2 diabetes, heterogeneity of treatment response is rarely assessed. This study aims to demonstrate the heterogeneity effects of SGLT2i for preventing chronic kidney disease (CKD) via a causal machine learning model.

Methods: A retrospective cohort study was conducted. Data from 2,280 type 2 diabetes patients were used to develop a double/debiased machine learning (DDML) model for CKD prediction; a conventional treatment effect (TE) model was also constructed for comparison. Model performance was evaluated using area under the receiver operating characteristic curve, and calibration plot. Subsequently, average treatment effects (ATEs) and conditional average treatment effects (CATEs) of SGLT2i versus non-SGLT2i were estimated and compared among models.

Results: A neural network (NN) was selected for the DDML framework for CKD prediction. Although the conventional TE model demonstrated slightly better discrimination and calibration, the DDML-NN produced comparable results. The ATEs (95% confidence interval) estimated by DDML-NN and conventional TE models were -0.0399 (-0.0600, -0.0199) and - 0.0424 (- 0.0616, - 0.0232), respectively. Corresponding CATE ranges were -0.3945 to 0.2495 and - 0.6968 to 0.0026. Patients gaining higher benefits from SGLT2i trended to be older, male, and overweight, compared to who were benefited less from SGLT2i.

Conclusion: The protective effect of SGLT2i against CKD was indicated, albeit heterogeneous. Counterfactual models including DDML-NN and conventional TE models demonstrated similar performance. However, data from a larger cohort are still required to draw a firm conclusion.