Development of a machine learning-based prediction model for postoperative delirium in frail elderly patients...

In frail older adults, the incidence of postoperative delirium (POD) is markedly increased, leading to greater morbidity, prolonged length of stay, and higher healthcare costs. An accurate POD prediction model can direct preventive strategies and improve patient outcomes. Employing advanced machine-learning techniques, this study develops a POD prediction model using comprehensive preoperative and intraoperative data.

We enrolled 2,089 frail patients aged ≥ 65 years undergoing general anesthesia for noncardiac surgery at Fuyang People's Hospital between February 2023 and February 2025. Thirty-eight baseline, anesthetic, and laboratory variables were extracted; missing data were handled by multiple imputation using chained equations (MICE). The dataset was randomly split 7:3 into training and validation sets. After feature selection with Boruta and LASSO, eight machine-learning models-logistic regression, random forest, support-vector classifier, XGBoost, artificial neural network, naïve Bayes, k-nearest neighbors, and decision tree-were trained and compared, with ROC-AUC as the primary metric, accompanied by accuracy, precision, recall, and F1 score. Model interpretability was achieved using SHAP analysis for the best-performing algorithm.

Among 2,089 frail elderly patients, the incidence of POD was 16.52%. After Boruta and LASSO identified 15 key predictors, the XGBoost model achieved an AUC of 0.813, outperforming the other seven algorithms. SHAP analysis identified MMSE score, Charlson Comorbidity Index, and age as the strongest predictors. External validation demonstrated high clinical utility on decision-curve analysis, with an ROC-derived sensitivity of 0.813 and specificity of 0.793, confirming robust performance without overfitting.

This study presents a robust XGBoost-based model for predicting postoperative delirium in frail elderly patients undergoing noncardiac surgery, demonstrating the potential of machine learning for clinical risk stratification. With its balanced performance and high accuracy, the model enables clinicians to identify high-risk patients and initiate timely interventions. Future work should focus on integration into clinical workflows and further external validation.