Automated machine learning-based model for the prediction of delirium in patients after surgery for degenerative spinal disease

Objective This study used machine learning algorithms to identify critical variables and predict postoperative delirium (POD) in patients with degenerative spinal disease. Methods We included 663 patients who underwent surgery for degenerative spinal disease and received general anesthesia. The LASSO method was used to screen essential features associated with POD. Clinical characteristics, preoperative laboratory parameters, and intraoperative variables were reviewed and were used to construct nine machine learning models including a training set and validation set (80% of participants), and were then evaluated in the rest of the study sample (20% of participants). The area under the receiver-operating characteristic curve (AUROC) and Brier scores were used to compare the prediction performances of different models. The eXtreme Gradient Boosting algorithms (XGBOOST) model was used to predict POD. The SHapley Additive exPlanations (SHAP) package was used to interpret the XGBOOST model. Data of 49 patients were prospectively collected for model validation. Results The XGBOOST model outperformed the other classifier models in the training set (area under the curve [AUC]: 92.8%, 95% confidence interval [CI]: 90.7%-95.0%), validation set (AUC: 87.0%, 95% CI: 80.7%-93.3%). This model also achieved the lowest Brier Score. Twelve vital variables, including age, serum albumin, the admission-to-surgery time interval, C-reactive protein level, hypertension, intraoperative blood loss, intraoperative minimum blood pressure, cardiovascular-cerebrovascular disease, smoking, alcohol consumption, pulmonary disease, and admission-intraoperative maximum blood pressure difference, were selected. The XGBOOST model performed well in the prospective cohort (accuracy: 85.71%). Conclusion A machine learning model and a web predictor for delirium after surgery for the degenerative spinal disease were successfully developed to demonstrate the extent of POD risk during the perioperative period, which could guide appropriate preventive measures for high-risk patients.