Abstract 4365487: A Novel Classification of Heart Failure Derived from the Nationwide JROADHF Cohort Using Unsupervised Machine Learning

Introduction/Background: Heart failure (HF) care is still steered almost solely by left-ventricular ejection fraction (LVEF), yet outcomes remain sub-optimal, suggesting that the HFrEF/HFmrEF/HFpEF scheme is too crude for precision therapy. Research Question/Hypothesis: We hypothesised that unsupervised machine learning (ML) applied to routinely collected variables could yield a more informative HF stratification than the conventional LVEF-based framework. Goals/Aims: To develop an ML-derived multidimensional HF classification, test its prognostic performance, and compare it with existing LVEF categories. Methods/Approach: We analysed 13,238 patients in the nationwide JROAD-HF registry in Japan. Forty-six clinical, laboratory, and echocardiographic variables entered a latent-class model; the cohort was randomly split 1:1 into discovery and internal-validation sets to assess robustness. (Figure 1) Clinical endpoint defined as a composite of 5-year cardiovascular death and heart failure readmission. Results: The model identified three phenogroups by minimum Bayesian Information Criteria. Phenogroup 1 “Advanced Low-Output HF” (predominantly male, age 73.6 ± 12 years) showed severe systolic dysfunction (LVEF 35.5 %), renal impairment, and heavy burdens of ischemic heart disease (52 %) and cardiomyopathy (29%). Phenogroup 2 “Early Afterload-Mismatch HF” (predominantly male, age 68.8 ± 12 years) was the youngest, markedly hypertensive, with intermediate LVEF (41.9%) and diverse etiologies. Phenogroup 3 “Elderly HFpEF-like HF” (predominantly female, age 84.5 ± 7 years) was lean and anemic, with preserved LVEF 53.5%, atrial fibrillation (45%), and valvular disease (42%). The four most discriminative variables were age, total bilirubin, LV diastolic diameter, and serum hemoglobin rather than LVEF and yielded an adjusted Rand index of 0.94 in the validation set, indicating excellent reproducibility. Five-year Kaplan–Meier curves showed the lowest mortality in Phenogroup 2 and the highest in Phenogroup 1 (log-rank p < 0.001), whereas LVEF categories showed no significant separation (p = 0.267) (Figure 2 A and B). Conclusions: ML-driven phenomapping of routine hospital data delineated three distinct HF groups that outperformed the traditional LVEF classification in prognostic discrimination.