Prognostic models for patients suffering a heart failure with a preserved ejection fraction: a systematic review

Comprehensive Summary

This systematic review evaluates prognostic models developed for patients with heart failure (HF) with preserved left ventricular ejection fraction (HFpEF). Researchers searched numerous databases, using key terms such as “heart failure,” “prognosis,”and “ROC curve,” to find articles about predictive models for patients with HFpEF published from February 1, 2023 to June 1, 2023. Articles that met eligible criteria were assessed with the Prediction Model Risk of Bias Assessment Tool (PROBAST) to group studies into groups of low, high, or unclear bias. Post-screening, sixteen articles remained, including a total of 39 prognostic models. Studies varied in geographic location, HFpEF definition, setting where models were developed, sample size, and outcomes. Common outcomes included all-cause mortality and rehospitalization, while common predictors included age, NT-pro BNP, and ejection fraction <60%. Out of all studies, only one assessed multiple biomarkers, despite their importance in HF prognostics. Overall, all studies were classified as high or unclear bias, due to high risk in the participant domain (retrospective datasets unrepresentative of greater populations) and/or analysis domain (not meeting the events per variable ≥20 criteria, weak screening methods, and/or mishandling variables). Researchers emphasize the importance of developing a study with a low risk of bias and evaluating critical metrics such as differentiation and calibration to accurately and wholly assess model performance.

Outcomes and Implications

Prognostic studies such as those evaluated in this systematic review usually provide a positive outlook on future ability to predict disease progression and identify higher-risk patients. However, this review sheds light on substandard quality studies, which only serves to slow the clinical implementation of possibly life-saving prognostic techniques. It is vital that these studies take due diligence to reduce prognostic model bias, include relevant predictors (such as biomarkers), perform external validations of machine learning models, and compare a model’s performance to modern prognosis techniques. These measures must be taken before clinical implementation can begin.