Machine learning models for predicting postoperative delirium after noncardiac surgery: A comparative study.

BackgroundPostoperative delirium is a frequent and serious complication after noncardiac surgery, linked to increased morbidity, prolonged hospitalization, and long-term cognitive decline. Although several prediction models have demonstrated good discriminative ability in external validation, challenges remain regarding implementation across clinical settings and model interpretability. This study compared three machine learning models-eXtreme Gradient Boosting, logistic regression, and support vector machine-for early postoperative delirium prediction.MethodsA retrospective cohort of 143 adults undergoing elective noncardiac surgery was analyzed (incidence of postoperative delirium = 15.4%). Data regarding 11 perioperative variables, including age, American Society of Anesthesiologists class, Mini-Mental State Examination score, surgery duration, and lowest intraoperative mean arterial pressure, were collected. Data were split in an 80:20 ratio into training and validation sets. Performance was assessed using area under the receiver operating characteristic curve, sensitivity, specificity, accuracy, and calibration. Shapley Additive Explanations analysis evaluated feature contributions.ResultsAge, Mini-Mental State Examination score, hemoglobin, surgery duration, opioid dose, lowest mean arterial pressure, blood loss, and American Society of Anesthesiologists class were significant predictors. eXtreme Gradient Boosting achieved the best validation performance (area under the receiver operating characteristic curve = 0.852; 95% confidence interval = 0.781-0.923), outperforming logistic regression (0.715) and support vector machine (0.698), with good calibration (Hosmer-Lemeshow,