Machine learning–based prognostic models for mortality prediction after unplanned hospitalization in cancer patients.

11122 Background: Unplanned hospitalizations in oncology patients often indicate acute clinical deterioration and are associated with high mortality. Prognostic assessment at admission remains a challenge, particularly in resource-limited settings. Machine learning (ML) methods can integrate clinical and laboratory features to generate individualized mortality predictions, although the data is scarce. We aimed to develop and validate ML-based models to predict in-hospital, 30-day, and 90-day all-cause mortality in adult cancer patients with unplanned hospitalizations. Methods: We retrospectively analyzed patients with cancer who were admitted for unplanned hospitalizations at Hacettepe University Oncology Hospital between January 2018 and August 2024. Class imbalance was addressed using SMOTE, and ANOVA-based feature selection retained the top 20 predictors. Nine ML algorithms were tested, including SVM, kNN, decision trees, neural networks, and ensemble models. Model performance was assessed via 5-fold cross-validation using accuracy, precision, recall, F1-score, and AUC. Results: A total of 1646 cancer patients with unplanned hospitalizations were included. The most common cancers were gastrointestinal malignancies (32.8%), lung cancer (17.4%), and breast cancer (9.8%). For in-hospital mortality prediction, Random Forest achieved the best performance on the original dataset (AUC-ROC: 0.89; accuracy: 84.6%). After SMOTE, SVM yielded the superior results (AUC-ROC: 0.98; accuracy: 93.6%). For 30-day mortality prediction, Random Forest was optimal pre-SMOTE (AUC-ROC:0.85; accuracy:83.6%), whereas LightGBM reached the highest post-SMOTE performance (AUC-ROC:0.97; accuracy:91.8). For 90-day mortality prediction, XGBoost demonstrated the best performance on the original dataset (AUC-ROC: 0.84; accuracy: 80.4%). Following the application of SMOTE, LightGBM outperformed the other models (AUC-ROC:0.90; accuracy:82.3%). Conclusions: This first study focusing on predicting in-hospital mortality in cancer patients with unplanned admissions shows that ML models based on routine admission data can reliably predict early- and long-term mortality. Incorporating such models into workflows may support early risk stratification, shared decision-making, and individualized care planning.