PREDICTING AND INTERPRETING ISCHEMIC STROKE USING A HYBRID STACKING ENSEMBLE MODEL (XGBOOST-LSTM WITH SHAP EXPLAINABILITY)

Ischemic stroke is a major cause of death and disability, and its early and accurate prediction is required for effective clini-cal intervention. Conventional machine learning models are robust but are not interpretable, which is of particular con-cern in clinical decision-making. A new hybrid stacking en-semble model is introduced in this paper, which combines Long Short-Term Memory (LSTM) networks and extreme Gra-dient Boosting (XGBoost) to improve the accuracy of predic-tion. Logistic regression meta-learner predicts finally by ag-gregating probabilistic outputs of the two base models. For explaining model prediction and determining important risk factors, Shapley Additive explanations(SHAP) is employed. Experiments on an actual ischemic stroke dataset demonstrate that the hybrid model outperforms the LSTM and XGBoost models alone in terms of accuracy and stability. With im-proved categorization performance, the stacking ensemble provides clinicians with an informative decision-support tool. Additionally, SHAP analysis reveals key variables such as age, BMI, and glucose, making the model more interpretable. With early diagnosis and aggressive medical interventions, these results indicate that the hybrid model can be a powerful tool for stroke risk prediction.