Stroke Prediction on Healthcare Data Using SMOTE and Explainable Machine Learning

The early detection and prevention of stroke stands as a vital healthcare priority because this condition leads to major disabilities and deaths worldwide. This study introduces an efficient machine learning system which predicts stroke risk through analysis of clinical and demographic information. The dataset contained an extreme class imbalance because strokes occurred in only 4.87 % of patients. The Synthetic Minority Over-sampling Technique (SMOTE) was used to balance the dataset until it reached a 50 % -50 % distribution. The research compared four machine learning algorithms through implementation: Lo-gistic Regression, Random Forest, XGBoost and CatBoost. The CatBoost classifier produced the best results among all models with an accuracy rate of 95.58 % and precision rate of 96.82 % and recall rate of 94.23 % and Fl-score of 95.51 %. The study employed explainable machine learning approaches to reveal how the model made decisions which improved both interpretation and clinical decision support. The proposed model demonstrates better performance than baseline methods and shows potential to assist clinicians with stroke risk assessment.