Optimising Stroke Recurrence Prediction Using Minimal Clinical Features and Machine Learning Models

Stroke recurrence remains one of the most devastating challenges in managing cerebrovascular disease, adding to disability, mortality, and rising healthcare costs worldwide. Being able to predict recurrence early could mean the difference between timely intervention and irreversible outcomes. In this study, we explored whether machine learning models - Logistic Regression, Random Forest, and XGBoost - could predict recurrence risk using only a small set of routine clinical features. Preprocessing involved managing missing values, scaling variables, and applying SMOTE to balance the classes without distorting real patient patterns. Models were evaluated across accuracy, precision, recall, F1 Score, and AUC-ROC, with greater weight placed on recall and F1 given the clinical need to minimize missed recurrences. Random Forest delivered the strongest results, achieving an accuracy of 92.39%, a recall of 94.05%, an F1 Score of 92.56%, and an AUC-ROC of 97.04%. These findings suggest that even simple, carefully designed predictive models could offer real clinical value, particularly in healthcare environments where rich data resources are limited and early warnings could make a critical difference for patient care.