Risk stratification and shap-derived mortality phenotypes in AMI and heart failure with breast/prostate cancer: NIS 2018–2022
V Godishala, V Polana, S Tanna, S Karipineni, N Patel, D Nanavaty, S Singh, S ReddyAbstract
Background
Cardio-oncology admissions represent a clinically complex population with heterogeneous drivers of in-hospital mortality. Although demographic and comorbidity factors are known contributors, it remains unclear whether distinct, explainable mortality “phenotypes” can be derived using admission-available features and contemporary machine learning approaches in nationally representative data.
Methods
Using NIS 2018–2022, we identified 27,975 admissions for acute myocardial infarction (AMI) and heart failure with concurrent breast or prostate cancer. The primary outcome was in-hospital mortality (6.0%). We used a temporally stratified design: 2018–2021 admissions for training (n=22,534) and 2022 for external testing (n=5,441). An XGBoost classifier incorporated demographics, hospital characteristics, cancer type, cardiac condition, Elixhauser comorbidity index, metastatic status, weekend admission, transfer-in status, and early diagnosis codes. Survey discharge weights were applied in training and evaluation. Performance was assessed using weighted AUC, average precision, Brier score, and decile-based calibration. SHAP quantified feature contributions. Mortality phenotypes were identified by clustering SHAP explanation vectors from the top 30 features, with the number of clusters chosen by silhouette analysis.
Results
Among 27,975 hospitalizations, overall mortality was 6.04%. In the 2022 external validation cohort (n=5,441), the model showed strong discrimination (AUC 0.829; average precision 0.316) and good calibration (Brier 0.054). Observed mortality increased across predicted-risk deciles, with close agreement in the highest-risk bin (predicted 0.309 vs observed 0.339). SHAP identified age, acute respiratory failure with hypoxia, AMI/ischemic heart disease, metastatic disease, female sex, payer status, COVID-19, and palliative care encounter codes as key predictors. Silhouette analysis supported five clusters (K=5). Weighted mortality ranged from 0.0% to 13.8%, defining distinct phenotypes: a respiratory failure–dominant high-risk group (13.8% mortality; predicted 14.7%); a COVID-associated group (9.1%; 9.4%); an elderly group (mean age 83.7; 7.4%; 7.2%); an ischemic/AMI-driven group (5.7%; 3.9%); and a metastatic/palliative-enriched group (0.0%; 1.8%). Predicted risk tracked observed mortality across clusters, supporting coherence and clinical interpretability.
Conclusions
An admission-only, temporally validated, explainable ML model achieved high discrimination for in-hospital mortality and identified five distinct SHAP-derived mortality phenotypes. These phenotypes reflect differing mechanistic drivers of risk (respiratory failure, ischemic cardiac pathology, metastatic burden, and systemic illness). They may support targeted risk stratification and pathway-specific care strategies in cardio-oncology populations.