Binary Classification Models for Stroke Outcome Prediction

Abstract

Stroke is found to be a leading cause of mortality and long-term disability worldwide, forcing effective predictive models to identify at-risk individuals and optimise treatment plans. In this study, we evaluate the performance of various machine learning (ML) algorithms in predicting stroke-related mortality. Five binary classification models—Logistic Regression (LR), Random Forest (RF), Gradient Boosting Machines (XGBoost), Support Vector Machine (SVM), and Neural Networks (MLPClassifier)-were applied to a dataset containing clinical and demographic features of stroke patients registered by the neurology department of the Clinical Centre of Montenegro. Each model was trained and evaluated using standard classification metrics: accuracy, precision, recall, and F1-score. Also, the importance of the feature was analysed to find the key predictors of stroke mortality across different models. The research shows the Random Forest and XGBoost performance over simpler models, proposing superior accuracy and interpretability. By analysing how precision, recall, and accuracy changes across a range of classification thresholds, we gained deeper insight into the model’s reliability under different clinical conditions. This analysis revealed clear trade-offs: lower thresholds improve recall (reducing the risk of missed death predictions), while higher thresholds enhance precision (minimising false positives). The findings support the selection of threshold values tailored to specific clinical priorities, such as early warning, balanced risk assessment, or high-confidence decision-making.