AI-delirium guard: Predictive modeling of postoperative delirium in elderly surgical patients

In older patients, postoperative delirium (POD) is a major complication that can result in greater morbidity, longer hospital stays, and higher healthcare expenses. 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.

We examined information from the National Surgical Quality Improvement Program (NSQIP), which included 17,000 patients who were over 65 and undergoing different types of surgery. 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. A range of machine learning approaches were assessed such as Decision Trees and Random Forests. Based on the greatest Area Under the Curve (AUC) from Receiver Operating Characteristic (ROC) analysis, the final model was chosen. Hyperparameter tuning was performed using GridSearchCV, optimizing parameters like max_depth, min_child_weight, and gamma for XGBoost model.

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

This study proposes a strong XGBoost-based model to predict POD in older surgical patients, demonstrating the potential of Machine Learning (ML) in clinical risk assessment. With the help of the model's balanced performance indicators and high accuracy, physicians may identify high-risk patients and promptly execute interventions in clinical settings. Subsequent investigations ought to concentrate on integration into clinical workflows and external validation.