In-hospital cardiac arrest (IHCA) from non-cardiac causes is a life-threatening condition characterized by high mortality and complex etiology. Despite advances in critical care monitoring, effective early warning systems remain lacking due to challenges in balancing sensitivity and specificity within imbalanced clinical data and the interpretability of machine learning models remains poorly defined.

We developed and evaluated six machine learning algorithms using data from 43,618 ICU patients in the MIMIC-IV database. Model performance was assessed through multiple metrics, and SHapley Additive exPlanations (SHAP) analysis was employed for model interpretability. Decision curve analysis was performed to evaluate clinical utility.

Distinct performance trade-offs were identified across models. XGBoost achieved optimal discriminative ability (AUC = 0.730, 95% CI: 0.695–0.763), while CatBoost favored sensitivity (0.907) to meet clinical guidelines. SHAP analysis identified anion gap (AG), white blood cell count (WBC), red cell distribution width (RDW), Glasgow Coma Scale (GCS), platelet, partial thromboplastin time (PTT), and red blood cell count (RBC) as key predictors, with elevated inflammatory markers and reduced neurological parameters consistently associated with increased risk.

Gradient boosting algorithms demonstrate crucial utility in predicting non-cardiac IHCA. The identified interpretable biomarkers align with established pathophysiological mechanisms and may reflect downstream manifestations of genome instability pathways, representing potential targets for early clinical intervention.