Development and External Validation of an Interpretable Machine Learning Framework for Predicting Pneumothorax-Associated Acute Kidney Injury: A Multicenter Retrospective Study

Background/Objectives: Intensive Care Unit (ICU) patients with pneumothorax face an elevated risk of developing Acute Kidney Injury (AKI) due to compromised hemodynamics and increased intrathoracic pressure. Early identification is crucial but challenging. This study aimed to develop and externally validate an interpretable machine learning (ML) framework to predict pneumothorax-associated AKI. Methods: This multicenter retrospective study utilized data from the MIMIC-IV database (development), alongside a temporal validation cohort (MIMIC-III) and an independent external validation cohort (eICU). A tri-algorithm intersection strategy—comprising Boruta, Least Absolute Shrinkage and Selection Operator (LASSO), and Recursive Feature Elimination (RFE)—was applied to extract optimal predictors. We systematically evaluated nine supervised ML algorithms. Shapley Additive exPlanations (SHAP) and Restricted Cubic Spline (RCS) analyses were integrated to unveil decision-making mechanics and non-linear dynamics. The optimal model was deployed as a web-based dynamic nomogram. Results: The hybrid feature selection strategy identified a parsimonious consensus of 7 core predictors (BUN, SOFA score, CKD, PEEP, Heart Failure, Albumin, and Age). Following comprehensive evaluation, the Logistic Regression model demonstrated favorable discriminative performance [Area Under the Curve (AUC) = 0.839 (95% CI: 0.786–0.891), Sensitivity = 81.5%, Specificity = 76.0%] and external generalizability (eICU AUC = 0.869; MIMIC-III AUC = 0.854). SHAP analysis delineated the individual contribution of each feature. RCS analysis revealed significant non-linear, dose–response relationships, highlighting an exponential AKI risk escalation driven by elevated BUN and higher PEEP levels. Decision Curve Analysis (DCA) suggested the potential net clinical benefit of the model across all validation cohorts. Conclusions: We developed a transparent, externally validated ML framework for predicting pneumothorax-associated AKI. The resulting web-based nomogram provides intensive care physicians with a practical, data-driven bedside tool to assist in personalized risk stratification. However, given the substantial calibration drift observed in the external cohort, local recalibration is essential prior to its use outside the MIMIC-derived population, and prospective cohort validation remains necessary before routine clinical implementation.