Machine Learning Model Uses Routine Lab Tests to Support Heart Failure Classification

Accurately distinguishing heart failure phenotypes is important for guiding imaging and treatment decisions in hospital care, but phenotyping usually depends on echocardiography, which may not be immediately available for every patient. Researchers have now developed a laboratory-based machine learning approach that uses routine test results to identify heart failure with reduced ejection fraction (HFrEF). By flagging patients more likely to have HFrEF, the model could support triage and help prioritize echocardiography when imaging resources are limited.

Investigators at the Eighth Affiliated Hospital of Guangxi Medical University developed and internally evaluated a model that draws on 13 routine laboratory indicators. The analysis included 1,480 hospitalized patients with chronic heart failure, comprising 377 with HFrEF and 1,103 with mildly reduced or preserved ejection fraction. Six algorithms were trained and assessed, including random forest, extreme gradient boosting (XGBoost), and logistic regression.

The HFrEF group showed higher levels of pro–B‑type natriuretic peptide, blood urea nitrogen, total and direct bilirubin, gamma‑glutamyl transferase, hematocrit, and uric acid than patients with mildly reduced or preserved ejection fraction. In an independent test set, random forest and XGBoost achieved areas under the receiver operating characteristic curve (AUROC) of 0.789. Logistic regression performed comparably with an area of 0.784 and demonstrated strong calibration with an intercept of 0.017 and slope of 1.069.

Discrimination was similar across models, with pairwise bootstrap comparisons showing no significant differences. The lowest Brier score of 0.152 was observed with random forest. At the default probability threshold, sensitivity was limited; adjusting the random‑forest threshold to 0.15 increased sensitivity to 0.912 and the negative predictive value to 0.935, supporting a rule‑out strategy that steers echocardiography toward higher‑risk patients.

Model explainability using Shapley Additive Explanations (SHAP) identified pro–B‑type natriuretic peptide as the dominant predictor. The findings were published online on Aug. 9 in Frontiers in Cardiovascular Medicine. The authors note that external and prospective validation will be needed before any clinical deployment.

“The model should be regarded as an adjunctive triage tool pending external and prospective validation, not as a substitute for echocardiographic phenotyping,” the authors stated.