Machine learning for stroke in heart failure with reduced ejection fraction but without atrial fibrillation: A post-hoc analysis of the WARCEF trial

Abstract

Background: The prediction of ischaemic stroke in patients with heart failure with reduced ejection fraction (HFrEF) but without atrial fibrillation (AF) remains challenging. Our aim was to evaluate the performance of machine learning (ML) in identifying the development of ischaemic stroke in this population.

Methods: We performed a post-hoc analysis of the WARCEF trial, only including patients without a history of AF. We evaluated the performance of 9 ML models for identifying incident stroke using metrics including area under the curve (AUC) and decision curve analysis. The importance of each feature used in the model was ranked by SAPley Additive exPlanations (SHAP) values.

Results: We included 2213 patients with HFrEF but without AF (mean age 58 ± 11 years; 80% male). Of these, 74 (3.3%) had an ischaemic stroke in sinus rhythm during a mean follow-up of 3.3 ± 1.8 years. Out of the 29 patient-demographics variables, 12 were selected for the ML training. Almost all ML models demonstrated high AUC values, outperforming the CHA₂DS₂-VASc score (AUC: 0.643, 95% confidence interval [CI]: 0.512-0.767). The Support Vector Machine (SVM) and XGBoost models achieved the highest AUCs, with 0.874 (95% CI: 0.769-0.959) and 0.873 (95% CI: 0.783-0.953), respectively. The SVM and LightGBM consistently provided significant net clinical benefits. Key features consistently identified across these models were creatinine clearance (CrCl), blood urea nitrogen (BUN) and warfarin use.

Conclusions: Machine-learning models can be useful in identifying incident ischaemic strokes in patients with HFrEF but without AF. CrCl, BUN and warfarin use were the key features.

Keywords: heart failure with reduced ejection fraction; machine learning; stroke.

FIGURE 1

The comparison of the receiver operating characteristic curves for 9 models. The shaded area around each curve indicates the 95% CI. CI, confidence interval; DT, decision tree; GBM, gradient boosting machine; KNN, K‐nearest neighbours; LR, logistic regression; MLP, multi‐layer perceptron; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine.

FIGURE 2

The comparison of the decision curve analysis for 9 models. DCA, decision curve analysis; DT, decision tree; GBM, gradient boosting machine; KNN, K‐nearest neighbours; LR, logistic regression; MLP, multi‐layer perceptron; RF, random forest; SVM, support vector machine.

FIGURE 3

Feature rankings according to SHAP values in the Support Vector Machine model. Each feature is organised according to SHAP values. BMI, body mass index; BUN, blood urea nitrogen; CrCl, creatinine clearance; DBP, diastolic blood pressure; DM, diabetes mellitus; DT, decision tree; HR, heart rate; KNN, K‐nearest neighbours; LR, logistic regression; LVEF, left ventricular ejection fraction; MLP, multi‐layer perceptron; SHAP, SHapley Additive exPlanations.

FIGURE 4

Feature rankings according to SHAP values in the XGBoost model. Each feature is organised according to SHAP values. BMI, body mass index; BUN, blood urea nitrogen; CrCl, creatinine clearance; DBP, diastolic blood pressure; DM, diabetes mellitus; HR, heart rate; LVEF, left ventricular ejection fraction; SHAP, SHapley Additive exPlanations.

FIGURE 5

Feature rankings according to SHAP values in the Random Forest model. Each feature is organised according to SHAP values. BMI, body mass index; BUN, blood urea nitrogen; CrCl, creatinine clearance; DBP, diastolic blood pressure; DM, diabetes mellitus; HR, heart rate; LVEF, left ventricular ejection fraction; RF, random forest; SHAP, SHapley Additive exPlanations.