Tehran University of Medical Sciences Archives of Anesthesiology and Critical Care 2423-5849 0 0 2026 11 07 1692 1692 Retti Kartika Bakti Department of Anesthesiology and Reanimation, Faculty of Medicine, Universitas Airlangga, Surabaya, East Java, Indonesia. & Department of Anesthesiology and Reanimation, Dr. Soetomo General Academic Hospital, Surabaya, East Java, Indonesia. Arie Utariani Department of Anesthesiology and Reanimation, Faculty of Medicine, Universitas Airlangga, Surabaya, East Java, Indonesia. & Department of Anesthesiology and Reanimation, Dr. Soetomo General Academic Hospital, Surabaya, East Java, Indonesia Nancy Rehatta Universitas Kristen Petra, Surabaya, East Java, Indonesia Adhi Wibawa Institut Teknologi Sepuluh Nopember, Surabaya, East Java, Indonesia Kohar Santoso Department of Anesthesiology and Reanimation, Faculty of Medicine, Universitas Airlangga, Surabaya, East Java, Indonesia. & Department of Anesthesiology and Reanimation, Dr. Soetomo General Academic Hospital, Surabaya, East Java, Indonesia. Lucky Andriyanto Department of Anesthesiology and Reanimation, Faculty of Medicine, Universitas Airlangga, Surabaya, East Java, Indonesia. & Department of Anesthesiology and Reanimation, Dr. Soetomo General Academic Hospital, Surabaya, East Java, Indonesia. Alief Wikarta Institut Teknologi Sepuluh Nopember, Surabaya, East Java, Indonesia 2026 07 18 2026 08 14 Background: This retrospective pilot study explored individual and ensemble machine learning models for predicting 30-day mortality in children with sepsis undergoing laparotomy. Methods: A retrospective analysis was conducted for 60 children diagnosed with sepsis who received laparotomy at a hospital in Indonesia between 2020 and 2024. K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest, and Extreme Gradient Boosting (XGBoost) were evaluated before and after tuning. To address class imbalance inherent in mortality datasets, Synthetic Minority Oversampling Technique (SMOTE) was applied. Model performance was assessed using accuracy, class-specific F1-score, and area under the receiver operating characteristic curve (AUC). Ensemble strategies, including soft voting and stacking classifiers, were additionally employed to optimize predictive performance. Results: The results showed that XGBoost and random forest demonstrated the highest accuracy among individual models. However, ensemble methods outperformed all single-model approaches. The stacking classifier achieved 100% accuracy on a small test set of 12 patients, while the soft voting ensemble demonstrated stable and consistent performance across metrics. The stacking classifier combining Random Forest and XGBoost, with logistic regression as the meta-learner and passthrough enabled, yielded an AUC of 0.70, indicating fair discriminative ability. Exploratory evaluation on an external dataset of six previously unused patients yielded a reported AUC of 0.62. Conclusion: The findings are exploratory. The small sample size, few deaths, modest reported discrimination, and lack of independent validation prevent claims of clinical readiness. Larger multicenter and prospective validation is needed before clinical use. https://aacc.tums.ac.ir/index.php/aacc/article/view/1692