Machine learning identification of risk factors for ischaemic stroke in patients with heart failure with reduced ejection fraction but without atrial fibrillation: A report from 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 and risk stratification based clinical factors alone has limited discrimination. Our aim was to evaluate the performance of machine learning (ML) to identify risk factors for ischaemic stroke in such HFrEF patients. Methods We performed a post-hoc analysis of the WARCEF trial, only including patients without prior history of AF, and censoring those who developed incident AF during follow-up. We evaluated the performance of 9 ML models to identify risk factors for incident stroke, using metrics including area under the curve (AUC), precision, specificity, and decision curve analysis. The importance of each feature in constructing the model was quantified using SHapley Additive exPlanations (SHAP) values. Results We included 2,213 patients with HFrEF but without AF (mean age 59±12 years; 79% male). Of these, 74 (3.3%) had an ischaemic stroke during mean follow-up of 3.5±1.8 years. Out of the 29 patient-demographics variables, 12 were selected for the ML training. All ML models demonstrated high AUC values, outperforming the CHA2DS2-VASc score (Figure 1). The logistic regression model exhibited a slight deviation from this trend. The XGBoost model achieved an AUC of 0.86 (95% confidence interval [CI]: 0.72-0.99), while the Random Forest model showed an AUC of 0.85 (95% CI: 0.74-0.96). Similarly, the Support Vector Machine and Light Gradient Boosting Machine models recorded AUCs of 0.85 (95% CI: 0.72-0.97) and 0.83 (95% CI: 0.70-0.96), respectively. Each of these models attained precision and specificity scores of 1.0, indicating high performance. These models consistently provided significant net clinical benefits. Key risk factors for stroke identified across these models were creatinine clearance and blood urea nitrogen. The use of warfarin appears to be protective. Conclusions Machine-learning models can identify risk factors for ischaemic stroke in patients with HFrEF but without AF. Worsening renal failure was predictive of an incident stroke, while warfarin use was protective against stroke (Figure 2).Figure 1Figure 2