Inpatient national mortality analysis in hospitalized patients with cancer diagnoses using machine learning.

Abstract

e13686 Background: In-hospital mortality among hospitalized patients with cancer diagnoses remains a major clinical and health-system burden. National data describing the prevalence of cancer diagnoses among hospitalized patients and associated inpatient mortality patterns are needed to better characterize high-risk populations. Methods: We performed a cross-sectional analysis of adult hospitalizations in the National Inpatient Sample (NIS) from 2016–2021 that included at least one ICD-10 cancer diagnosis code (C00–C96), totaling 3,310,634 admissions. A tree-based gradient boosting machine learning model was developed to predict in-hospital mortality. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Results: The most prevalent cancer diagnoses in the dataset were secondary bone cancer (11.3%), secondary liver cancer (10.8%), prostate cancer (7.0%), lung cancer (5.5%), and secondary brain cancer (5.4%). Overall in-hospital mortality was 5.7%. The machine learning model demonstrated strong discrimination (ROC-AUC 0.859; accuracy 94.6%). Among frequently observed cancer diagnoses, ICD-10 codes corresponding to acute leukemia were most commonly present among patients who died during hospitalization, with an in-hospital mortality rate of 13.7%. Other cancer diagnoses with high inpatient mortality included disseminated cancer, secondary adrenal cancer, and primary liver cancer. SHAP analysis identified age, acute kidney injury, sepsis, non-elective admission, and payer status as major contributors to mortality prediction. Conclusions: In this national inpatient analysis of hospitalizations with cancer diagnoses, secondary malignancies were highly prevalent, while acute leukemia ICD-10 diagnoses were most commonly observed among patients who experienced in-hospital death. Machine learning with SHAP-based interpretability identifies clinically relevant inpatient mortality patterns that may assist early risk stratification during hospitalization.