Predicting major adverse cardiovascular and cerebrovascular events in chronic heart failure: a machine learning study

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Not specified in abstract

• Primary Outcome: Major adverse cardiovascular and cerebrovascular events (MACCE)

• Nnt Or Sensitivity: CatBoost model demonstrated best performance; HRR1 and VE/VCO2 slope identified as key predictors

• Confidence Interval: Not reported

Clinical Application

CPET is available in major Australian cardiac centres but requires specialised equipment and trained personnel CPET is covered under Medicare for specific indications; findings could inform risk stratification in Australian heart failure clinics aligned with NHFA-CSANZ guidelines Patients with chronic heart failure undergoing cardiopulmonary exercise testing

Abstract

BACKGROUND: Heart failure (HF) is a clinical syndrome characterized by impaired cardiac diastolic and systolic function due to structural or functional damage to the myocardium. HF represents the end-stage manifestation of many cardiac diseases. Therefore, early identification of high-risk patients is crucial. This study aims to utilize machine learning (ML) methods to develop and validate a model to predict major adverse cardiovascular and cerebrovascular events (MACCE) in patients with HF and identify its key predictive features. METHODS: This study is a retrospective cohort study. We enrolled a total of 271 patients, who were divided into training and testing sets. Baseline data, including cardiopulmonary exercise testing (CPET) parameters and laboratory tests, were collected for all participants. Based on the presence or absence of MACCE during follow-up, they were categorized into a No-event group and MACCE group. We developed seven ML models to predict the incidence of MACCE in patients with chronic heart failure (CHF) using CPET parameters. The predictive performance of these models was systematically compared, and model interpretability was evaluated using Shapley Additive exPlanations (SHAP). Subsequently, retaining only those with HF with preserved ejection fraction (HFpEF) for a sensitivity analysis. Additionally, a subgroup analysis was conducted between No-event group and Worsening HF (WHF) group. RESULTS: We used Boruta feature selection, four important predictive features were identified. Among the ML models constructed with these features, the Categorical Boosting (CatBoost) model demonstrated the best performance. SHAP analysis was applied to interpret the optimal model, revealing that lower values of heart rate recovery at 1 min (HRR1), as well as a higher carbon dioxide ventilation equivalent slope (VE/VCO2 slope), were associated with higher SHAP values-indicating greater importance in predicting adverse outcomes. CONCLUSION: HRR1 and VE/VCO2 slope can serve as independent predictors for the incidence of MACCE in patients with CHF. Both impaired HRR1 and an elevated VE/VCO2 slope serve as significant independent predictors of MACCE in the CHF population.HRR1 is an independent predictor of WHF in patients with CHF.