DEVELOPMENT AND PROSPECTIVE MULTICENTER VALIDATION OF AN EXPLAINABLE ARTIFICIAL INTELLIGENCE–DRIVEN PRECISION PROGNOSTIC MODEL FOR EARLY PREDICTION OF MORTALITY AND MULTIORGAN FAILURE IN CRITICALLY ILL ADULTS
ABSTRACT
Background: Early prediction of mortality and multiorgan failure remains a major challenge in critical care medicine. Conventional prognostic scoring systems have limited ability to capture the complex and dynamic interactions underlying critical illness. This study aimed to develop and prospectively validate an explainable artificial intelligence (AI)-driven precision prognostic model for early prediction of mortality and multiorgan failure in critically ill patients.
Materials and Methods: In this prospective multicenter cohort study, 3,012 adult ICU patients from five tertiary-care hospitals were enrolled. Clinical, physiological, laboratory, and treatment-related variables obtained within the first 24 hours of ICU admission were used to develop machine-learning models. The cohort was randomly divided into derivation (70%) and validation (30%) datasets. Multiple algorithms were evaluated, including logistic regression, random forest, artificial neural networks, and extreme gradient boosting (XGBoost). Model performance was assessed using AUROC, sensitivity, specificity, calibration metrics, and SHAP-based explainability analysis.
Result: Among 3,012 patients, in-hospital mortality occurred in 561 (18.6%) patients, while 729 (24.2%) developed multiorgan failure. XGBoost demonstrated the best performance, achieving an AUROC of 0.946 in the derivation cohort and 0.934 in the validation cohort for mortality prediction. The model significantly outperformed APACHE II (AUROC 0.806) and SOFA (AUROC 0.831) scores. For multiorgan failure prediction, the model achieved AUROCs of 0.923 and 0.912 in the derivation and validation cohorts, respectively. SHAP analysis identified serum lactate, SOFA score, vasopressor requirement, age, and renal dysfunction as the most influential predictors.
Conclusion: The explainable AI-driven prognostic model demonstrated excellent accuracy, calibration, and generalizability for early prediction of mortality and multiorgan failure in critically ill patients. This approach may enhance precision risk stratification and support timely clinical decision-making in intensive care settings.
Keywords: Artificial intelligence, Machine learning, Critical care, Intensive care unit, Mortality prediction, Multiorgan failure, Explainable AI, XGBoost, Precision medicine, Prognostic model.
Background: Early prediction of mortality and multiorgan failure remains a major challenge in critical care medicine. Conventional prognostic scoring systems have limited ability to capture the complex and dynamic interactions underlying critical illness. This study aimed to develop and prospectively validate an explainable artificial intelligence (AI)-driven precision prognostic model for early prediction of mortality and multiorgan failure in critically ill patients.
Materials and Methods: In this prospective multicenter cohort study, 3,012 adult ICU patients from five tertiary-care hospitals were enrolled. Clinical, physiological, laboratory, and treatment-related variables obtained within the first 24 hours of ICU admission were used to develop machine-learning models. The cohort was randomly divided into derivation (70%) and validation (30%) datasets. Multiple algorithms were evaluated, including logistic regression, random forest, artificial neural networks, and extreme gradient boosting (XGBoost). Model performance was assessed using AUROC, sensitivity, specificity, calibration metrics, and SHAP-based explainability analysis.
Result: Among 3,012 patients, in-hospital mortality occurred in 561 (18.6%) patients, while 729 (24.2%) developed multiorgan failure. XGBoost demonstrated the best performance, achieving an AUROC of 0.946 in the derivation cohort and 0.934 in the validation cohort for mortality prediction. The model significantly outperformed APACHE II (AUROC 0.806) and SOFA (AUROC 0.831) scores. For multiorgan failure prediction, the model achieved AUROCs of 0.923 and 0.912 in the derivation and validation cohorts, respectively. SHAP analysis identified serum lactate, SOFA score, vasopressor requirement, age, and renal dysfunction as the most influential predictors.
Conclusion: The explainable AI-driven prognostic model demonstrated excellent accuracy, calibration, and generalizability for early prediction of mortality and multiorgan failure in critically ill patients. This approach may enhance precision risk stratification and support timely clinical decision-making in intensive care settings.
Keywords: Artificial intelligence, Machine learning, Critical care, Intensive care unit, Mortality prediction, Multiorgan failure, Explainable AI, XGBoost, Precision medicine, Prognostic model.
View PDF Back