Publication

Abstract : Accurate and interpretable treatment evaluation is essential in Intensive Care Units (ICUs), where clinical decisions must account for rapidly evolving physiological states. Traditional machine learning models achieve high predictive accuracy but remain limited by their reliance on correlations, making them unsuitable for estimating true treatment effects. This study investigates whether Causal Artificial Intelligence (Causal AI) can complement predictive models by providing robust and clinically meaningful treatment insights. Using the eICU Collaborative Research Database, we analyzed demographics, vital signs, laboratory values, and treatment timelines to evaluate the effect of early antibiotic administration on 30-day mortality. Logistic Regression, Random Forest, and XGBoost were trained for outcome prediction, with ensemble models achieving strong performance (AUC ≈ 0.99) compared to Logistic Regression (AUC = 0.68). Causal estimators—including Propensity Score Matching (PSM) and Inverse Probability Weighting (IPW)—indicated a mortality reduction of 5–17% among patients receiving early antibiotics. These results show that while predictive models excel in discrimination, causal inference provides actionable, interpretable, and clinically aligned treatment-effect estimates. The combined framework underscores the potential of Causal AI to support trustworthy and decision-centric care in critical settings. We also perform leakage-aware feature selection and explicitly state causal assumptions and limitations of observational ICU data.