Development of machine learning models for predicting delirium in elderly patients after thoracic surgeries

Abstract

Postoperative delirium (POD) is a frequent neurological complication following anesthesia and surgery, which is associated with increased comorbidities and mortality. The current study developed and internally validated a machine learning model for predicting POD in elderly patients undergoing thoracic surgeries based on machine learning (ML) algorithms. Elderly patients who underwent elective segmentectomy, lobectomy, or esophagectomy with general anesthesia from January 2015 to December 2022 were electronically retrieved. Clinical features, preoperative laboratory values, and intraoperative parameters were collected. The least absolute shrinkage and selection operator (LASSO) regression was used to screen critical features associated with POD. Nine ML algorithms were used to construct prediction models in the training dataset (80% of participants), and model performances were evaluated in the validation dataset (20% of participants). Prediction performances of the nine ML models were also compared using the area under the receiver operating characteristic curve (AUC). Model calibration was assessed using the Brier scores. Finally, the best model was determined and was interpreted using the Shapley Additive exPlanations (SHAP). Three thousand nine hundred sixty-seven elderly patients were enrolled, and 277 of them developed POD. Preoperative white blood cell (WBC), preoperative fasting plasma glucose (FPG), intraoperative hypotension, arrhythmia, combined nerve block anesthesia, hypertension, intraoperative vasoactive drugs, current smoker, blood transfusion, glucocorticoids usage, ASA classification, and video-assisted thoracoscopy were identified as prediction variables. The XGBoost model exhibited the best predictive performance (AUC: 0.936, 95% confidence interval (CI): 0.932–0.941). The XGBoost model also presented the lowest Brier scores (0.110) in terms of model calibration. The three predictive variables with the most influence on the model were preoperative WBC, preoperative FPG and intra-operative hypotension. The current study developed a ML model for POD prediction for elderly patients undergoing thoracic surgeries. If validated, the model could potentially serve as a complementary tool to aid POD risk stratification.