Advancing Heart Failure Classification Beyond Left Ventricular Ejection Fraction: Unsupervised Machine Learning Insights From the Japanese Nationwide Acute Decompensated Heart Failure Cohort: JROADHF

Background Heart failure (HF) classification based on left ventricular ejection fraction (LVEF) inadequately captures HF heterogeneity, limiting phenotype‐specific therapies. While machine learning (ML) has been applied to HF with preserved ejection fraction, comprehensive approaches across the full LVEF spectrum remain scarce. This study aimed to identify HF phenogroups across the entire LVEF spectrum using unsupervised ML, moving beyond the LVEF‐centric and fragmented understanding of HF. Methods We analyzed 13 238 patients with acute HF from a nationwide, multicenter Japanese cohort (JROADHF [Japanese Registry of Acute Decompensated Heart Failure]). Unsupervised latent class analysis identified phenogroups from 46 clinically important admission variables. The primary end point was the 5‐year composite of cardiovascular death or HF readmission. Results Three ML‐based phenogroups emerged: (1) “HF with congestion and lower output” (n=2638), (2) “HF with afterload mismatch” (n=3462), and (3) “HFpEF‐like HF” (n=7138). ML‐derived groups stratified risk for the primary end point ( P <0.001), whereas LVEF‐based classification did not ( P =0.267). Compared with phenogroup 3, phenogroup 1 had a higher age‐adjusted risk by Cox regression model (hazard ratio [HR], 1.60 [95% CI, 1.49–1.71]) and phenogroup 2 had a lower risk (HR, 0.65 [95% CI, 0.60–0.70]). Notably, variable‐importance analysis by discriminative power for phenogroup assignment indicated that total bilirubin, left ventricular diastolic diameter, and hemoglobin rather than LVEF principally drove classification. Conclusions Unsupervised ML identified novel HF phenogroups with distinct 5‐year outcomes. This framework moves beyond an LVEF‐centric paradigm, deepens understanding of HF pathophysiology, and may facilitate phenotype‐specific therapeutic development.