A machine learning risk score predicts mortality across the spectrum of left ventricular ejection fraction

Aims: Heart failure (HF) guideline recommendations categorize patients according to left ventricular ejection (LVEF). Mortality risk, however, varies considerably within each category and the likelihood of death in an individual patient is often uncertain. Accurate assessment of mortality risk is an important component in the decision-making process for many therapies. In this report, we assess the accuracy of MARKER-HF, a recently described machine learning-based risk score, in predicting mortality of patients in the three guideline-defined HF categories and its ability to distinguish risk of death for patients within each category.

Methods and results: MARKER-HF was used to calculate mortality risk in a hospital-based cohort of 4064 patients categorized into groups with reduced, mid-range, or preserved LVEF. MARKER-HF was substantially more accurate than LVEF in predicting mortality and was highly accurate in all three HF categories, with c-statistics ranging between 0.83 to 0.89. Moreover, MARKER-HF accurately discriminated between patients at high, intermediate and low levels of mortality risk within each of the three categories of HF used by guidelines.

Conclusions: MARKER-HF accurately predicts mortality in patients within the three categories of HF used in guidelines for management recommendations and it discriminates between magnitude of risk of patients in each category. MARKER-HF mortality risk prediction should be helpful to providers in making recommendations regarding the advisability of therapies designed to mitigate this risk, particularly when they are costly or associated with adverse events, and for patients and their families in making future plans.