Comprehensive Summary
This retrospective study developed and externally validated an artificial intelligence (AI) model to accurately predict mortality and prolonged intensive care unit (ICU) stay in postoperative critically ill patients. Data from 6,029 postoperative ICU patients across two centers were analyzed using multiple machine-learning algorithms. A soft-voting ensemble model was developed to integrate the individual algorithms to generate final predictions. Following external validation, SHapley Additive exPlanations (SHAP) analysis identified the most influential clinical variables driving the model’s predictions. Internally, the ensemble model outperformed individual algorithms, achieving AUROCs of 0.8812 (mortality) and 0.7944 (prolonged ICU stay) with accuracies of 0.91 and 0.94, respectively. External validation demonstrated consistent performance, with AUROCs of 0.8330 (mortality) and 0.7376 (prolonged ICU stay) with corresponding accuracies of 0.92 and 0.90, respectively. Across models, SHAP analysis highlighted emergency surgery, serum osmolality, lactate levels, and diastolic blood pressure as the most influential predictors.
Outcomes and Implications
Existing prediction models for postoperative critically ill patients remain limited by suboptimal accuracy and generalizability. The application of AI offers a promising avenue to enhance predictive precision, as demonstrated by this study. The developed AI model outperformed traditional scoring systems in predicting both mortality and prolonged ICU stay among postoperative ICU patients. Despite limitations such as its retrospective design, two-center scope, and lack of calibration analysis, the model shows potential to guide clinical decision-making. Future multicenter prospective trials are needed to confirm the model’s accuracy and generalizability across diverse clinical settings.