An interpretable XGBoost and SHAP framework for predicting 3-month all-cause mortality in lung adenocarcinoma with liver metastasis

Patients with lung adenocarcinoma (LUAD) and synchronous liver metastasis have a high risk of early death. We evaluated whether a model restricted to information available at diagnosis could discriminate all-cause death within 3 months. We analyzed 7,395 patients diagnosed between 2010 and 2015, including zero-month deaths. The primary model used a prespecified set of diagnosis-time demographic, tumor, stage, and metastasis variables; chemotherapy, radiation, surgery, and grade were excluded. A fixed stratified split yielded 5,176 training and 2,219 validation records. Preprocessing, hyperparameter selection, threshold selection, and model fitting were restricted to the training data. Seven classifiers and a stacking ensemble were benchmarked, with chemotherapy and cancer-specific mortality evaluated only in sensitivity analyses. In the internal holdout cohort, baseline XGBoost achieved an AUC of 0.636 and an AUPRC of 0.651; logistic regression achieved an AUC of 0.645 and an AUPRC of 0.664. DeLong’s paired test showed no statistically significant AUC difference (XGBoost minus logistic regression, -0.008; p = 0.052). A sensitivity XGBoost model that included chemotherapy had an AUC of 0.815, but this association was not interpreted as a treatment effect. Diagnosis-time baseline variables showed modest discrimination for 3-month all-cause mortality. XGBoost did not demonstrate superior discrimination over logistic regression in this dataset. The model and the accompanying interactive risk-stratification prototype require temporal, geographic, and prospective impact validation before any clinical use.