OBJECTIVE:
To systematically compare the performance of nine commonly used machine learning algorithms in predicting ten-year mortality risk in ischemic stroke based on a unified clinical cohort, providing empirical evidence for algorithm selection.

METHODS:
Utilizing retrospective cohort data from 627 first-ever ischemic stroke patients in a Chinese island county. Predictors were selected using Lasso regression, Boruta algorithm, and recursive feature elimination. Nine algorithms were compared: Logistic Regression, XGBoost, LightGBM, Random Forest, AdaBoost, Decision Tree, GBDT, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). Data were split 8:2 into training and validation sets. Evaluation metrics included AUC, accuracy, sensitivity, specificity, calibration curves, decision curve analysis, and precision-recall curves.

RESULTS:
Eleven core predictors were finally selected, including age at onset, mRS score, discharge outcome, serum chloride (Cl), serum creatinine (Cr), etc. On the validation set, the AdaBoost model demonstrated the best robustness (AUC=0.800), along with good calibration and clinical net benefit. Random Forest achieved the highest AUC on the training set (0.918) but its performance decreased on the validation set (AUC=0.792), suggesting overfitting risk. XGBoost and Logistic Regression also showed good performance. SHAP interpretability analysis revealed that age at onset and mRS score were the most important predictors across all models.

CONCLUSIONS:
For ten-year mortality risk prediction in ischemic stroke, the AdaBoost algorithm demonstrated the best robustness and overall performance on this dataset. Different algorithms offered complementary perspectives on the same feature set. Age at onset and neurological deficit (mRS) were the most consistently strong core risk factors.