Background <p>Stroke is a leading cause of mortality and disability worldwide, creating a critical need for accurate prediction tools to support early clinical decision making. However, mortality prediction is often hindered by class imbalance, as death events represent a minority of cases in most clinical datasets. Various machine learning (ML) approaches and imbalance-handling techniques, including sampling methods, cost-sensitive learning, and threshold adjustment, have been proposed to address this challenge. This study aimed to compare these strategies and identify an optimal ML framework for predicting in-hospital mortality among patients with stroke.</p> Methods <p>Data from 2653 patients with stroke were analyzed. Five machine learning algorithms (XGBoost, RF, DNN, SVM, and LR) were evaluated using nested stratified 10-fold cross-validation. Class imbalance was addressed using threshold adjustment, cost-sensitive learning, and sampling methods (ENN, OSS, SMOTE, SMOTE-ENN, and SVM-SMOTE). Model performance was assessed using Accuracy, F1-score, G-mean, MCC, AUROC, and AUPRC. Statistical comparisons were performed using Friedman and post-hoc tests, and model interpretability was evaluated using SHAP.</p> Results <p>Among the baseline models, XGBoost achieved the highest discrimination performance, with an AUROC of 0.898 ± 0.014 and an AUPRC of 0.656 ± 0.076. Support Vector Machine and Logistic Regression achieved the highest G-mean values, indicating a better balance between sensitivity and specificity. Threshold adjustment based on the G-mean criterion improved the G-mean of XGBoost from 0.661 to 0.804 and increased the F1-score from 0.551 to 0.624, while cost-sensitive learning achieved a comparable G-mean of 0.806. Among the sampling strategies, ENN produced the highest F1-score (0.615 ± 0.026), whereas SMOTE-ENN achieved the highest G-mean (0.799 ± 0.032). Although all sampling methods improved minority-class detection, statistical analysis showed no significant differences among sampling strategies (Friedman χ² = 9.44, <i>p</i> = 0.051). SHAP analysis identified consciousness status, awareness level, patient arrival condition, Glasgow Coma Scale category, respiratory complications, length of hospital stay, and age as the most influential predictors of in-hospital mortality.</p> Conclusion <p>XGBoost achieved the best overall performance for predicting in-hospital mortality among patients with stroke. Threshold adjustment and cost-sensitive learning substantially improved minority-class detection, whereas ENN and SMOTE-ENN provided the most favorable sampling results. Overall, algorithm selection and threshold optimization had a greater impact on predictive performance than the specific sampling strategy. These findings support the potential use of machine-learning models for early risk stratification in stroke care.</p>