From guidelines to precision care: a personalised approach to heart failure management using heterogeneous treatment effects

Background Guideline-directed medical therapy (GDMT) for heart failure (HF) with reduced ejection fraction (HFrEF) is based on large randomised controlled trials in well-defined patient populations. Therefore, the GDMT recommendations are based on a one-size-fits-all assumption and fails to optimise individual outcomes due to patient heterogeneity e.g. in baseline risk, comorbidities, renal function, and haemodynamic tolerance. Moreover, the pivotal trials underlying the guideline recommendations do not consider dose-response effects, thereby ignoring individual variation in treatment response. Purpose The purpose of this study is to enable precision medicine in HFrEF management by developing a causal inference machine learning framework that provides patient-specific medication titration recommendations. Methods Patient-level data HFrEF/HFmrEF (HF with mildly reduced ejection fraction) patients in four studies (23,982 records) were harmonised, and doses were mapped to reflect ESC target-dose percentage. For each titratable class ACEi/ARB, beta-blockers, MRAs and loop diuretics, we identified consecutive visits where only the index class changes (up-titrate vs maintain) while other therapies were stable. Individualised 90-day absolute risk differences for all-cause death, HF hospitalisation or New York Heart Association (NYHA) functional class worsening were estimated using causal forests with covariate-balancing propensity weights (Figure 1). Validation included improved covariate balance between “up-titrate” and “maintain” groups, cross-validated outcome models (AUCs of 0.731–0.767) showed consistent discrimination, and recovery of known heterogeneous effects in synthetic data was successful. Results Individualised up-titration effects spanned benefit, near-zero effect and potential harm across drug classes, indicating heterogeneity in predicted response to up-titration. As an example, we focused on beta-blockers and ACEi/ARB to illustrate drug-to-drug titration prioritisation. Beta-blocker and ACEi/ARB effects were weakly correlated (Figure 1), yielding four responses quadrants (benefit from both, benefit from neither, beta-blockers only, ACEi/ARB only). Quadrant phenotyping revealed distinct profiles (Figure 2): patients favouring both agents had fewer comorbidities, those favouring neither had more, while mixed groups differed in renal function, IL6, and comorbidities. The model recommends up-titration more often than dosing decisions in actual data: beta-blockers 77.2% vs 14.1% and ACEi/ARB 50.5% vs 11.5%. Bootstrap-enabled confidence intervals for patient-level individual treatment effects were wide, indicating limited precision and supporting a hypothesis-generating interpretation of our dosing recommendations. Conclusions The framework suggests potential HF treatment sequencing through individualised up-titration recommendations and reveals clinically plausible patient heterogeneity but requires prospective validation.figure_1_flowchart_causalfigure_2_BB_RAS_quadrants