OBJECTIVE:
To explore the value of Stacking ensemble learning and Random Survival Forest (RSF) models in enhancing the performance of ischemic stroke mortality risk prediction and providing dynamic risk assessment.

METHODS:
Based on the same cohort of 627 patients and 11 predictors. First, three base learners (XGBoost, AdaBoost, KNN) with relatively good and complementary performance were selected. A two-layer Stacking ensemble model was constructed using 5-fold cross-validation to generate meta-features and Logistic Regression as the meta-learner. Second, a time-to-event prediction model based on Random Survival Forest was built, evaluating its discriminative ability (C-index) and time-dependent AUC. All final models were evaluated on an independent test set.

RESULTS:
The Stacking ensemble model achieved the best performance on the test set with an AUC of 0.834, significantly outperforming any single base learner, and provided the highest clinical net benefit across a wide range of threshold probabilities. The Random Survival Forest model showed good discriminative ability with a C-index of 0.736. Time-dependent analysis indicated that this model performed excellently in short-term (1-year) prediction (test set AUC=0.865) and maintained stable performance in medium- to long-term prediction.

CONCLUSIONS:
The Stacking ensemble strategy effectively integrates the strengths of different algorithms, significantly improving the predictive model's discriminative power and robustness. The Random Survival Forest model provides an effective tool for assessing dynamic long-term mortality risk, excelling particularly in short-term prognosis prediction. Their combination offers a methodological paradigm for constructing high-performance, robust ischemic stroke prognosis prediction tools with a temporal dimension.