Causal Explainability for ICU Mortality Prediction Using Electronic Health Record Time-Series Data

Predicting Intensive Care Unit (ICU) mortality from Electronic Health Record (EHR) time-series data remains a complex challenge due to the high dimensionality and limited interpretability of machine learning models. This study introduces Causal SHAP, a novel explainability framework that integrates causal graph discovery with SHAP-based feature attribution to produce causally consistent and clinically interpretable explanations for mortality prediction. Using the MIMIC-IV dataset of adult ICU patients, multiple models including Random Forest, XGBoost, and Long Short-Term Memory (LSTM) networks were trained to predict in-hospital mortality. The proposed Causal SHAP adjusts SHAP values using causal weights derived from the Peter–Clark (PC) and Greedy Equivalence Search (GES) algorithms, ensuring that feature attributions reflect true physiological dependencies rather than mere correlations. Experimental results showed that the Causal SHAP-enhanced LSTM achieved an AUC of 0.880, preserving predictive accuracy while improving causal consistency by 22.6% and clinical agreement with expert judgment by 19.1%. Key causal pathways identified, such as Lactate → Vasopressor → Mortality and Creatinine → MAP → Mortality, correspond closely to established mechanisms of hemodynamic and renal dysfunction in critical care. These findings demonstrate that embedding causal reasoning into explainable AI enables more transparent, trustworthy, and clinically grounded predictions for decision support in ICU settings.