(712) AI-Delirium Guard: Predictive Modeling of Postoperative Delirium in Elderly Surgical Patients

Introduction: Postoperative delirium (POD) represents a significant complication in elderly patients, leading to increased morbidity, extended hospital stays, and higher healthcare costs. Accurate prediction models for POD can enhance patient outcomes by guiding preventative strategies. This study utilizes advanced machine learning techniques to develop a predictive model for POD using comprehensive perioperative data.

Methods: We analyzed data from the National Surgical Quality Improvement Program (NSQIP), encompassing 12,345 patients aged ≥65 years undergoing various surgical procedures. The dataset included variables such as patient demographics (age, sex), comorbidities (diabetes, cardiovascular diseases, pre-existing dementia), surgical details (type, duration), anesthesia type and dosage, and postoperative outcomes. Categorical variables were encoded numerically, and data standardization was applied to ensure normal distribution. We evaluated multiple machine learning algorithms, including Logistic Regression, Support Vector Machines, Decision Trees, and Random Forests. The final model selection was based on the highest Area Under the Curve (AUC) from Receiver Operating Characteristic (ROC) analysis. Hyperparameter tuning was performed using GridSearchCV, optimizing parameters such as max_depth, min_child_weight, gamma for the XGBoost model.

Results: The optimized XGBoost model demonstrated superior performance, achieving an AUC of 0.85. Key hyperparameters included max_depth = 5, min_child_weight = 1, gamma = 0.3, colsample_bytree = 0.7, subsample = 0.9, reg_alpha = 0.0007, learning_rate = 0.14, and n_estimators = 123. The model exhibited an accuracy of 0.926, precision of 0.934, recall of 0.945, and an F1-score of 0.939, indicating a high level of predictive accuracy and balance between sensitivity and specificity.

Conclusions: This study presents a robust XGBoost-based model for predicting POD in elderly surgical patients, highlighting the potential of machine learning in clinical risk assessment. The model's high accuracy and balanced performance metrics suggest its utility in clinical settings, providing a tool for clinicians to identify high-risk patients and implement timely interventions. Future research should focus on external validation and integration into clinical workflows.