Machine learning vs. traditional regression analysis for fluid overload prediction in the ICU

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Fluid overload is a common complication in intensive care units (ICUs), leading to adverse outcomes.
• Predicting fluid overload is difficult, and ICU medication use may influence its development.
• Machine learning (ML) offers potential advantages over traditional regression for predicting complex clinical outcomes.

Purpose of the Study:

• To compare the predictive performance of traditional regression techniques and ML models for identifying fluid overload in ICU patients.
• To identify clinically meaningful predictors of fluid overload, including patient and medication-related factors.

Main Methods:

• A retrospective observational cohort study of adult ICU patients (≥72 hours stay) with available fluid balance data.
• Development and comparison of traditional logistic regression and various supervised ML models (e.g., XGBoost) to predict fluid overload (≥10% weight gain).
• Evaluation of model performance using Area Under the Receiver Operating Characteristic (AUROC), Positive Predictive Value (PPV), and Negative Predictive Value (NPV).

Main Results:

• Out of 391 patients, 12.5% developed fluid overload.
• The XGBoost ML model achieved the highest performance (AUROC 0.78, PPV 0.27, NPV 0.94).
• XGBoost performance was comparable to the optimal traditional logistic regression model (AUROC 0.70, PPV 0.20, NPV 0.94).
• Feature importance highlighted illness severity scores and medication data as key predictors.

Conclusions:

• Both ML and traditional models demonstrated similar effectiveness in predicting ICU fluid overload in this cohort.
• Baseline severity of illness and the complexity of ICU medication regimens are significant predictors of fluid overload.
• ML approaches, particularly XGBoost, show promise for fluid overload prediction, with feature importance offering clinical insights.