Development and External Validation of an Interpretable Machine Learning Framework for Predicting...

Area of Science:

• Critical Care Medicine
• Nephrology
• Artificial Intelligence in Healthcare

Background:

• Intensive Care Unit (ICU) patients with pneumothorax have a higher risk of developing Acute Kidney Injury (AKI).
• Early AKI prediction in these patients is challenging due to hemodynamic instability and increased intrathoracic pressure.
• Timely identification of at-risk patients is crucial for intervention and improved outcomes.

Purpose of the Study:

• To develop and externally validate an interpretable machine learning (ML) framework for predicting pneumothorax-associated AKI.
• To identify key predictors of AKI in ICU patients with pneumothorax.
• To create a practical, data-driven tool for bedside risk stratification.

Main Methods:

• A multicenter retrospective study using MIMIC-IV for development and MIMIC-III/eICU for validation.
• A tri-algorithm intersection strategy (Boruta, LASSO, RFE) for optimal predictor selection.
• Evaluation of nine supervised ML algorithms, with SHAP and RCS analyses for interpretability.

Main Results:

• A parsimonious set of 7 core predictors identified: BUN, SOFA score, CKD, PEEP, Heart Failure, Albumin, and Age.
• Logistic Regression model achieved strong performance (AUC=0.839) with good external generalizability (eICU AUC=0.869, MIMIC-III AUC=0.854).
• SHAP and RCS analyses revealed key drivers and non-linear relationships, particularly elevated BUN and PEEP.

Conclusions:

• A transparent, externally validated ML framework for predicting pneumothorax-associated AKI was successfully developed.
• The web-based nomogram serves as a practical bedside tool for physicians to aid in personalized risk stratification.
• Local recalibration and prospective validation are recommended before routine clinical implementation due to observed calibration drift.