AHA Scientific Sessions 2026 Abstract · Nov 2, 2026

An X-Learner Causal Forest Meta-Analysis for Personalized SGLT2i/GLP-1 Selection in HFpEF

By Ghosh A, Khalid F, Arif Z et al.

Background: Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous syndrome. Standard meta-analyses pool trial data to generate a single Average Treatment Effect (ATE). SGLT2 inhibitors (DELIVER: dapagliflozin HR 0.82 [95% CI 0.73-0.92]; EMPEROR-Preserved: empagliflozin HR 0.79 [95% CI 0.69-0.90]) and GLP-1 receptor agonists (STEP-HFpEF: semaglutide, KCCQ improvement) show efficacy, but the ATE conceals profound phenotypic heterogeneity. Methods: We systematically searched PubMed, MEDLINE, EMBASE, and the Cochrane Library from inception to March 2026 to identify relevant studies and synthetic cohort of N=2,000 HFpEF patients was generated, calibrated to published baseline characteristics (mean age 71±10 years, BMI 30±6 kg/m ² , 57% with atrial fibrillation, 63% diabetic, mean LVEF 54±8%, mean eGFR 62 mL/min/1.73m ² ). Subgroup-specific potential outcomes were derived from published hazard ratios across 10 clinical phenotypic features. Two causal meta-learner frameworks were applied: (1) X-Learner with Gradient Boosting base models, and (2) Causal Forest Double Machine Learning (EconML), both estimating the Conditional Average Treatment Effect τ(x) = E[Y(1) – Y(0) ∣ X = x] at the individual patient level. Results: The Causal Forest yielded a mean CATE of –1.6 (SD 10.3), with the X-Learner confirming mean CATE –1.3 (SD 11.8), demonstrating substantial individual-level heterogeneity. SGLT2i was the preferred strategy in 40.8% of patients; GLP-1 RA in 59.2%. The forest plot revealed striking phenotypic divergence: patients with atrial fibrillation (CATE +3.0), non-obese BMI (CATE +6.0), and preserved renal function favoured SGLT2i; whereas obese patients (BMI ≥30, CATE –9.1) and those with diabetes (CATE –5.1) derived superior benefit from GLP-1 RA. Causal Forest feature importance identified BMI (0.711), atrial fibrillation (0.150), and HbA1c (0.110) as the dominant treatment-effect modifiers. A prescribing heatmap across the BMIхHbA1c phenotypic space delineated four actionable phenotypes: lean/euglycaemic (SGLT2i dominant), obese/diabetic (GLP-1 RA dominant), lean/diabetic (equipoise), and obese/euglycaemic (borderline). Conclusions: Causal meta-learner frameworks applied to HFpEF trial subgroup data reveal that the ATE is an inadequate summary statistic. Personalised prescribing algorithms, driven primarily by BMI, glycaemic status, and atrial fibrillation, can identify which HFpEF phenotype derives superior benefit from SGLT2i versus GLP-1 RA. These findings mandate prospective validation and reframe HFpEF pharmacotherapy toward precision medicine.

doi.org/10.1161/circ.154.suppl_1.9241Interventions

AgingResearch.News