Light Tree-Based Models for Mortality Prediction of ICU Patients

Numerous predictive models have been developed to assess in-hospital mortality risk of patients receiving intensive care, particularly those on mechanical ventilation. However, light tree-based models can accelerate clinical outcome predictions and enable timely interventions by physicians. An automated feature selection framework is proposed in this study, including Autofeat library, ReliefF and Fast CorrelationBased Filter (FCBF). The proposed framework was employed to eliminate most of the features before the prediction task. Treebased models - Bagging (BGG) and Extreme Gradient Boosting (XGB) - demonstrated comparable predictive accuracy using only 7 and 4 features for the imbalanced sample (S1.1) and balanced sample (S2.2), respectively, out of 67 original features. The BGG model, utilizing 7 features, achieved high predictive accuracy with an Area Under the Curve (AUC) of 0.89 (95 % confidence interval <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[CI]: 0.87-0.90)</tex>. In contrast, when applied to the S2.2 sample with only 4 features, both the XGB and BGG model yielded inferior results, as reflected by an AUC of 0.81 (95% CI: 0.79-0.83).