Advancing Stroke Prediction: Explainable Machine Learning Models and Data Augmentation Strategies

Stroke, a leading global disability cause, affects 12 million people annually, with 6.5 million dying. Preventing at least half of strokes requires public policy action and increased awareness. Early diagnosis is crucial and enhances patients' opportunities to survive a stroke. This research looks into how well machine learning (ML) models can predict when strokes will happen and how data augmentation affects those predictions.. Five ML models (Logistic Regression, Decision Tree, K-Nearest Neighbors, XGBoost, and Random Forest) are assessed and their performance is compared using an imbalanced and augmented dataset. Data augmentation techniques, including Random Oversampling, SMOTE, and Cluster Centroids, are employed to address class imbalance. The results demonstrate that some models achieve relatively high accuracy without data augmentation (XGBoost: 0.95). However, data augmentation techniques further improve performance across all models. Random Oversampling leads to the highest Accuracy for Random Forest (0.99), while SMOTE achieves the best Precision for Logistic Regression (1.00). The results found that age, occupation type, BMI, blood glucose level, and history of cardiovascular disease are risk factors for stroke. This study highlights the effectiveness of ML models, particularly when combined with data augmentation, for stroke prediction and emphasizes the importance of specific features in identifying stroke risk.