Machine Learning Prediction of Early In-Hospital Mortality in Lung Cancer Patients Using Administrative Data.

Abstract

This study develops and validates an interpretable machine learning model for predicting early in-hospital mortality (death within 7 days of admission) among lung cancer patients using diagnosis codes available at admission. We analyzed the 2016-2019 Maryland State Inpatient Database (HCUP SID), with patient history captured from 2014, identifying 57,625 adult lung cancer patients who died during hospitalization. Of these, 40,651 (70.5%) experienced early death. One hundred binary clinical features were derived from ICD-10-CM codes mapped to AHRQ Clinical Classifications Software Refined categories. XGBoost achieved an AUC of 0.787 (95% CI: 0.782- 0.792) at baseline, improving to 0.791 with engineered features, outperforming logistic regression (0.783), random forest (0.769), and traditional comorbidity indices (Charlson: 0.367; Elixhauser: 0.325). Notably, comorbidity indices performed worse than chance, suggesting chronic disease burden is inversely associated with rapid deterioration. Interpretability analyses using gain importance, permutation importance, and SHAP consistently identified acute metabolic derangements, electrolyte imbalances, malnutrition, and severe infections as dominant predictors. Risk stratification showed monotonic separation with observed early mortality ranging from 31.5% (lowest quintile) to 95.6% (highest quintile). These findings demonstrate that administrative data can support real-time risk stratification for palliative care and intensity-of-care decisions in acute oncology settings.