Distinguishing heart failure subtypes quickly has real consequences: patients with reduced ejection fraction respond to therapies — beta-blockers, ACE inhibitors, SGLT2 inhibitors — that carry little or no benefit for preserved ejection fraction phenotypes. When echocardiography is unavailable or delayed, clinicians are effectively flying blind. A machine-learning framework trained on standard blood draws could change that triage window meaningfully.
This single-center retrospective study enrolled 1,480 hospitalized chronic heart failure patients, of whom 377 had HFrEF and 1,103 had mid-range or preserved ejection fraction phenotypes. Using only the first blood draw within 24 hours of admission, a leakage-free pipeline applied recursive feature elimination to distill 13 routine laboratory indicators — including proBNP, hematocrit, total and direct bilirubin, blood urea nitrogen, beta-2 microglobulin, CRP, uric acid, globulin, GGT, mean corpuscular hemoglobin, LDL-C, and fibrinogen. Among six algorithms tested, random forest and XGBoost each achieved an AUC of 0.789 on the held-out test set, while logistic regression delivered a nearly identical AUC of 0.784 with superior calibration, suggesting the simpler model may be more deployable in clinical settings.
Several aspects deserve scrutiny. The 0.789 AUC, while clinically useful as a screening adjunct, falls short of replacing echocardiography and should be interpreted as a triage accelerator rather than a diagnostic substitute. The single-center design limits generalizability — laboratory reference ranges and patient demographics vary considerably across institutions and geographies. The retrospective framework also cannot demonstrate that earlier HFrEF identification actually improves outcomes; that requires prospective validation. On the positive side, the SHAP interpretability layer and decision-curve analysis add clinical credibility beyond raw discrimination metrics. The finding that a well-calibrated logistic regression matches ensemble model performance is practically significant: simpler models are easier to audit, validate, and integrate into electronic health record workflows. This work is incremental but directionally important, pointing toward a future where inexpensive blood panels routinely pre-stratify cardiac phenotypes before imaging slots become available.