Machine-Learning Model Uses Routine Blood Tests to Flag a Common Type of Heart Failure

A machine-learning model built from 13 routine laboratory tests can help distinguish heart failure with reduced ejection fraction (HFrEF) — a form of heart failure where the heart's main pumping chamber squeezes weakly — from other types of heart failure, according to a study published online August 9 in Frontiers in Cardiovascular Medicine.

Zhiping Meng of the Eighth Affiliated Hospital of Guangxi Medical University in Guigang, China, and colleagues analyzed 1,480 hospitalized patients with chronic heart failure — 377 with HFrEF and 1,103 with heart failure with mildly reduced or preserved ejection fraction. The team tested six algorithms using 13 routine indicators, including pro-B-type natriuretic peptide (proBNP), blood urea nitrogen, bilirubin, an enzyme called gamma-glutamyl transferase, hematocrit and uric acid, all of which were higher in the HFrEF group.

Random forest and XGBoost achieved the best performance, both scoring 0.789 on a scale called the area under the receiver operating characteristic curve, or AUC, which measures how well a model separates two groups, with logistic regression close behind at 0.784. proBNP was the single strongest predictor. At the models' default settings, sensitivity — the ability to correctly flag true HFrEF cases — was limited, but adjusting the random-forest model's threshold to 0.15 raised sensitivity to 0.912 and its negative predictive value to 0.935, meaning it could reliably rule out HFrEF and help decide which patients need an echocardiogram first.

The researchers said the model should be used as an adjunctive triage tool pending external and prospective validation, not as a substitute for echocardiographic phenotyping — the standard heart-imaging method for classifying heart failure types.