Interpretable Machine Learning Predicts Successful Weaning and ICU discharge alive Following Prone Positioning Ventilation in Moderate-to-Severe Acute Respiratory Distress Syndrome (ARDS) Secondary to Pneumonia.

OBJECTIVE: We developed and validated interpretable machine learning models predicting successful weaning and survival to ICU discharge in patients with moderate-to-severe acute respiratory distress syndrome (ARDS) secondary to pneumonia receiving prone position ventilation (PPV). METHODS: We included 131 patients with moderate-to-severe ARDS secondary to pneumonia undergoing PPV for internal, temporal, and external validation. Models were developed using least absolute shrinkage and selection operator (LASSO) regression. Performance was evaluated using the area under the curve (AUC), calibration, and decision curve analysis (DCA). Model interpretability was assessed using SHapley Additive exPlanations (SHAP). RESULTS: The support vector machine (SVM) model had the highest internal AUC (0.715 weaning; 0.807 ICU discharge). The weaning model was stable in temporal (AUC 0.841) and external (AUC 0.707) validation. The ICU discharge model had limited temporal/external stability (AUCs 0.657, 0.665) and was preliminary. Calibration was good for weaning and acceptable for ICU discharge. DCA suggested potential utility across broad thresholds. SHAP identified key predictors for weaning (septic shock, platelet count, PaO2/FiO2 ratio at PPV, neutrophil-to-lymphocyte ratio (NLR), and peak airway pressure at PPV) and ICU discharge (APACHE II, peak airway pressure at PPV, multidrug-resistant organism infection, and NLR). CONCLUSIONS: The weaning SVM model showed promising but exploratory performance and may serve as an adjunct for individualized risk stratification during the conventional ventilation/PPV phase, complementing rather than replacing clinical judgement. The ICU discharge model was less stable and should be interpreted cautiously. SHAP improved transparency and generated hypotheses, not causal evidence. Further multi-center prospective validation is needed.