Predicting survival in patients with severe heart failure: risk scores validation in the HELP-HF cohort

Background: Accurate selection of patients with severe heart failure (HF) who might benefit from advanced therapies is crucial. Several risk scores aimed at predicting the risk of mortality in patients with HF are available, but their performance in patients with severe HF has been poorly investigated. Methods: The risk of 1-year mortality was estimated in the cohort of the HELP-HF registry, including 1,149 patients with severe HF, according to the MAGGIC, 3-CHF, ADHF/NTproBNP, GWTG risk scores, as well as to the number of criteria of the 2018 HFA-ESC definition of advanced HF, the number of I NEED HELP markers, the INTERMACS profile, the number of frailty domain fulfilled and the frailty index. The accuracy of the scores was calculated with the area under the receiver operator characteristic curve (AUC) analysis. In addition, we tested the performance of different machine learning (ML)-based models to predict 1-year mortality. Results: At 1-year follow-up, 265 patients (23.1%) died. The prognostic accuracy was moderate for MAGGIC (AUC 0.71), 3C-HF (AUC 0.71), and ADHF/NTproBNP scores (AUC 0.70) and only modest for the other scores. All the scores lost accuracy in estimating the rate of 1-year mortality in patients at the highest risk. Support vector machine-based model had the best AUC among ML-based models, outperforming most of the tested risk scores. Conclusion: Most of the scores used to predict the risk of mortality in HF are only partially applicable to real-world patients with severe HF. MAGGIC and GWTG-HF scores showed the best discriminative ability but provided inaccurate estimate of the risk of 1-year mortality in patients at the highest risk. ML-based models might improve risk prediction in these patients.