An interpretable machine learning framework with clinical nomogram for predicting in-hospital mortality in acute ischemic stroke using high-granularity bedside data.

BACKGROUND: This multicenter study developed and validated an interpretable machine learning model integrating granular nursing and emergency department data collected within the first 24 hours to predict in-hospital mortality in acute ischemic stroke (AIS).

METHODS: We analyzed a retrospective cohort of 5,014 adult AIS patients from three tertiary academic centers (2019-2023). Centers A and B (n = 3,512) formed the development cohort; Center C (n = 1,502) served as the external validation cohort. Sixty-three predictors across seven domains were extracted from the initial 24 hours. A consensus feature selection approach combining LASSO, RFE-RF, filter methods, and XGBoost importance was employed, and six machine learning algorithms were evaluated using nested cross-validation, Bayesian optimization, SMOTE, and rigorous anti-leakage protocols. SHAP values and a logistic nomogram enhanced interpretability.

RESULTS: The best-performing CatBoost model with 22 RFE-RF features achieved AUC-ROC of 0.917 (95% CI 0.899-0.933) internally and 0.891 (95% CI 0.868-0.912) externally, significantly outperforming APACHE III, SOFA, OASIS, and GCS (ΔAUC 0.142-0.193, all p < 0.001). Calibration was excellent (slopes 0.934-0.977, Brier 0.098-0.114). A pre-specified 0-12 h sensitivity analysis confirmed predictive validity is not dependent on late-stage trajectories (external AUC 0.873). A 10-feature nomogram achieved external AUC 0.871 (ΔAUC -0.020 vs CatBoost) with superior DCA net benefit over all comparators.

CONCLUSIONS: Integration of bedside nursing assessments with emergency and laboratory data into an ensemble gradient-boosting model markedly improves early in-hospital mortality prediction in AIS compared with established ICU scores.