Development and Validation of a Machine-Learning Model for Predicting Prognosis in Critically Ill Patients Undergoing Major Surgery

Background Major surgery can result in elevated mortality rates, poorer prognoses, and extended hospital stays. This study sought to develop and validate an effective machine-learning model capable of accurately forecasting outcomes in critically ill patients who have undergone major surgery. Methods Using the publicly accessible Medical Information Mart for Intensive Care (MIMIC)-IV database, we developed and validated multiple machine-learning models to predict postoperative outcomes in critically ill patients who had at least an overnight ICU stay after major surgery. Seven predictive models were tested to forecast prognosis, with the highest-performing model selected based on its accuracy and the area under the receiver operating characteristic curve (AUC). An advanced model, eXtremely Gradient Boosting (XGBoost), was created using all variables, followed by a streamlined model built from 10 features chosen for their importance and clinical applicability. The performance of both models was assessed using Decision Curve Analysis (DCA), while survival analyses distinguished high- and low-risk groups within the internal validation sets. Results A cohort of 2,335 critically ill patients who had undergone major surgery were included in the MIMIC-IV cohorts. The full XGBoost model achieved an accuracy of 80.6% and an AUC of 0.828, indicating high predictive power. A more practical selection model with 10 features demonstrated a slightly lower AUC of 0.824 (95% CI: 0.762–0.886) but offered advantages in clinical usability. The ten key features were identified based on their SHapley Additive exPlanations (SHAP) values, which included the Charlson Comorbidity Index (CCI), Simplified Acute Physiology Score (SAPS)-II, Sequential Organ Failure Assessment (SOFA) score, mechanical ventilation, ARDS and sepsis complications, blood urea nitrogen (BUN) levels, estimated glomerular filtration rate (eGFR), respiratory rate, and marital status. Additionally, survival curves showed a clear distinction between high- and low-risk groups based on predictions from both the full and selection XGBoost models. Conclusions This study developed two XGBoost model variants that outperform other ICU predictive methods for forecasting the prognosis of patients undergoing major surgery. These models have the potential to assist healthcare providers in making more informed decisions, thereby improving clinical outcomes in the ICU setting.