Cervical cancer (CC) remains a significant cause of cancer mortality among women in low- and middle-income countries, particularly in regions with limited healthcare resources. This study aims to develop and evaluate machine learning (ML) models for hospital-level risk stratification of in-hospital mortality among patients hospitalized with CC in Mato Grosso, Brazil, between 2011 and 2023.

Hospitalization data for CC patients were obtained from the SUS Hospital Information System (SIH). The dataset included demographic, clinical, and hospital variables. Five ML algorithms: Logistic Regression, Random Forest, CatBoost, LightGBM, and XGBoost, were implemented. Model performance was evaluated using AUC-ROC, calibration (Brier score), and threshold-dependent measures (accuracy, sensitivity, specificity, precision, and F1-score). Model interpretability was examined using Shapley Additive Explanations (SHAP).

A total of 3,493 hospitalisations were analysed. Among the evaluated models, XGBoost achieved the best overall performance, with an AUC-ROC of 0.89, accuracy of 88%, and specificity of 95%, and demonstrated good calibration (Brier score 0.08). Tree-based models, including Random Forest and CatBoost, also showed high specificity (> 93%) and low Brier scores. SHAP analysis identified medical procedure type, hospitalization cost, and service complexity as the most influential predictors of in-hospital mortality.

ML models demonstrated robust performance in estimating in-hospital mortality risk among cervical cancer patients using administrative hospitalization data. By combining predictive performance with model interpretability, these approaches provide insights into clinical and system-level factors associated with mortality and may support hospital-level risk stratification, health system evaluation, and resource allocation in resource-limited settings.