Razi University

  Accurate prediction of mortality risk in patients undergoing hematopoietic stem cell tra  lantation (HSCT) plays a vital role in clinical decision-making. Although ensemble machine learning models have been widely employed in recent years to improve the accuracy of these predictions, most previous studies have modeled clinical outcomes as simple binary events, disregarding the time-to-event structure and censoring information. Furthermore, systematic and statistically rigorous comparisons of different ensemble strategies in this field remain limited. The primary objective of this study was to develop and evaluate an ensemble framework that integrates survival analysis, including the Kaplan–Meier method and Cox regression, with machine learning models, and to compare three ensemble strategies—simple averaging, weighted averaging, and stacking—using appropriate statistical tests.This study was conducted using the publicly available IN13-01 dataset from the Center for International Blood and Marrow Tra  lant Research (CIBMTR), comprising 1,673 patients who underwent HSCT. To prevent information leakage, 19 variables associated with post-tra  lant outcomes were excluded. A Cox proportional hazards regression model was fitted using ten clinical variables, and the resulting risk score, together with four interaction features, was added to the set of predictors used by the machine learning models, including logistic regression, random forest, and gradient boosting. Finally, the three ensemble strategies were evaluated on an independent test set comprising 25% of the data using the area under the receiver operating characteristic curve (AUC), F1-score, accuracy, sensitivity, and specificity. Pairwise comparisons of model performance were conducted using DeLong’s test.The Kaplan–Meier analysis indicated that conditioning regimen intensity, tra  lant type, performance status, age, and tra  lant period were significantly associated with overall survival (p < 0.01, HR = 1.147). In the multivariable Cox model, age (p < 0.001, HR = 0.827) and performance status (p < 0.001, HR = 0.827) were identified as the strongest independent predictors, and the model achieved a concordance index of C-index = 0.611. Furthermore, the Cox risk score was identified as the most important feature in the machine learning models, with a permutation importance of 86%, substantially exceeding that of the other features.Among the ensemble strategies, stacking achieved the highest performance, with an AUC of 0.6913, and demonstrated a reported statistically significant improvement over logistic regression (DeLong’s test, p = 0.17). However, the difference in performance between stacking and the best-performing base model, random forest (AUC = 0.689), was not statistically significant (p = 0.726).Overall, the findings of this study suggest that integrating survival analysis information with machine learning models through a Cox risk score is an effective approach for jointly leveraging time-to-event information and clinical characteristics to predict mortality risk. Nevertheless, increasing the complexity of ensemble strategies, from simple averaging to stacking, does not necessarily result in a statistically significant improvement over the best-performing base model. Therefore, the evaluation and selection of ensemble models should be based not only on numerical comparisons of performance metrics but also on appropriate statistical significance tests.