Compare SGLT2I versus non-SGLT2I users in type-2 diabetic mellitus patients on GLP-1 receptor agonist: A population-based and machine learning causal inference analysis

Zhiyao Luo, Oscar Hou In Chou, Zita Man Wai Ng, Cheuk To Chung, Jeffrey Shi Kai Chan, Raymond Chan, Lei Lü, Tingting Zhu, Bernard Man Yung Cheung, Tong Liu, Gary Tse, Jiandong ZhoumedRxiv 2023 · DOI

Introduction

Type-2 diabetes mellitus (T2DM) remains a primary driver of global cardiovascular morbidity. While sodium-glucose cotransporter-2 inhibitors (SGLT2I) and GLP-1 receptor agonists (GLP1a) have independently demonstrated significant cardioprotective benefits, clinical evidence regarding their combined use is limited. Most clinical trials evaluate these drug classes in isolation, leaving a critical evidence gap for clinicians managing high-risk patients already on GLP1a therapy. This study utilizes real-world data and machine learning-based causal inference to investigate whether adding an SGLT2I to a GLP1a regimen provides additive cardiovascular protection.

Methods

This retrospective population-based cohort study utilized the Clinical Data Analysis and Reporting System (CDARS) from the Hong Kong Hospital Authority, covering the period from January 1, 2015, to December 31, 2020. The cohort included 2,526 T2DM patients already receiving GLP1a therapy, categorized into SGLT2I users (n=1,968) and non-users (n=558).

To ensure robust results, the researchers employed a dual-analytical approach:

  1. Propensity Score Matching (PSM): Patients were matched at a 1:2 ratio based on demographics, comorbidities, and baseline medications to minimize selection bias.
  2. Machine Learning Causal Inference: Causal Random Forests (CRF) were applied to estimate the Conditional Average Treatment Effect, accounting for complex, non-linear interactions between patient variables.

The primary outcomes were new-onset myocardial infarction (MI), heart failure (HF), atrial fibrillation, and stroke/TIA. All-cause mortality was the secondary outcome.

Results

The study found that adding SGLT2I to GLP1a therapy significantly reduced the risk of major mechanical cardiovascular events. After adjusting for confounders, non-users of SGLT2I faced a nearly three-fold higher risk of myocardial infarction and a 2.5-fold higher risk of heart failure compared to combination users.

Graphical abstract showing study design and key hazard ratios for myocardial infarction and heart failure.

Key findings include:

  • Myocardial Infarction: Non-users had a Hazard Ratio (HR) of 2.91 (95% CI: 1.30–6.59).
  • Heart Failure: Non-users had an HR of 2.49 (95% CI: 1.22–5.08).
  • Clinical Impact: The Number Needed to Treat (NNT) was 49 to prevent one heart attack and 45 to prevent one case of heart failure.
  • Other Outcomes: While risks for atrial fibrillation (HR 1.52) and stroke (HR 1.72) were numerically higher in non-users, these did not reach statistical significance.

Subgroup analyses indicated that the benefits of SGLT2I were consistent across various demographics and comorbidities, including age, sex, and the presence of hypertension or hyperlipidaemia.

Discussion and Conclusion

These findings suggest a synergistic cardioprotective effect when SGLT2Is are added to GLP1a therapy. The protection appears most pronounced for "mechanical" cardiac issues (MI and HF) rather than "electrical" (atrial fibrillation) or "vascular" (stroke) issues in this population. Notably, the benefits remained consistent regardless of fasting glucose levels, suggesting that the mechanisms of protection extend beyond simple glycemic control.

While this observational study is limited by the lack of lifestyle data (e.g., BMI, smoking) and the potential for residual confounding, the alignment between traditional statistical models and machine learning causal inference provides strong evidence for the benefits of combination therapy. Future large-scale randomized controlled trials are warranted to confirm these real-world findings.

Summary updated

Abstract

Abstract Background Both sodium-glucose cotransporter-2 (SGLT2) inhibitors and GLP-1 receptor agonists (GLP1a) demonstrated benefits against cardiovascular diseases in type 2 diabetes (T2D). However, the effects of SGLT2I amongst patients already on GLP1a users remain unknown. Objective This real-world study compared the risks of cardiovascular diseases with and without exposure to SGLT2I amongst GLP1a users. Methods This was a retrospective population-based cohort study of patients with type-2 diabetes mellitus (T2DM) on GLP1a between 1st January 2015 and 31st December 2020 using a territory-wide registry from Hong Kong. The primary outcomes were new-onset myocardial infarction, atrial fibrillation, heart failure, and stroke/transient ischaemic attack (TIA). The secondary outcome was all-cause mortality. Propensity score matching (1:2 ratio) using the nearest neighbour search was performed. Multivariable Cox regression was used to identify significant associations. The machine learning causal inference analysis was used to estimate the treatment effect. Results This cohort included 2526 T2DM patients on GLP1a (median age: 52.5 years old [SD: 10.9]; 57.34 % males). The SGLT2I users and non-SGLT2I users consisted of 1968 patients and 558 patients, respectively. After matching, non-SGLT2I users were associated with high risks of myocardial infarction (Hazard ratio [HR]: 2.91; 95% Confidence Interval [CI]: 1.30-6.59) and heart failure (HR: 2.49; 95% CI: 1.22-5.08) compared to non-SGLT2I users after adjusting for demographics, comorbidities, medications, renal function, and glycaemic tests. However, non-SGLT2I users were not associated with the risks of atrial fibrillation (HR: 1.52; 95% CI: 0.65-3.53) and stroke/TIA (HR: 1.72; 95% CI: 0.70-4.24). The results remained consistent in the competing risk and the sensitivity analyses. Conclusions SGLT2I non-users was associated with higher risks of myocardial infarction and heart failure when compared to SGLT2I users after adjustments amongst T2DM patients on GLP1a. The result remained consistent in the machine learning causal inference analysis. Graphical abstract
BibTeX
@misc{luo2023compare,
  author = {Zhiyao Luo and Oscar Hou In Chou and Zita Man Wai Ng and Cheuk To Chung and Jeffrey Shi Kai Chan and Raymond Chan and Lei Lü and Tingting Zhu and Bernard Man Yung Cheung and Tong Liu and Gary Tse and Jiandong Zhou},
  title = {{Compare SGLT2I versus non-SGLT2I users in type-2 diabetic mellitus patients on GLP-1 receptor agonist: A population-based and machine learning causal inference analysis}},
  year = {2023},
  doi = {10.1101/2023.11.06.23298185},
  howpublished = {medRxiv},
}
Figure