Development and validation of Machine Learning model to predict refractory septic shock

This study developed and validated a Machine Learning (ML) model to predict refractory septic shock (RSS) in patients with sepsis admitted to a tertiary care center in India. Using an ambispective design, data from 1,008 adult ICU patients were used for model development and 102 for prospective validation. Demographic, clinical, laboratory, and imaging variables were analyzed through a structured three-tiered feature selection process, and Random Forest classifiers were trained on optimized feature sets. The best-performing model, incorporating 27 clinical and laboratory features, achieved an AUROC of 0.877 in the training cohort and 0.839 in prospective validation, demonstrating high accuracy, precision, and recall. Early identification of high-risk patients using this model can facilitate timely interventions and improve outcomes. The validated ML model shows strong predictive ability and interpretability for RSS, though multicenter studies are required to confirm its generalizability before widespread clinical implementation.