Machine Learning Algorithm for Predicting Postoperative Delirium in Elderly Patients After Thoracic Surgery

Development and Validation of a Prediction Model for Postoperative Delirium in Elderly Patients After Thoracic Surgery

Postoperative delirium (POD) is a common and severe complication in patients undergoing major surgery, especially in the elderly. POD has been proven to be associated with increased morbidity and mortality, institutionalization, and high healthcare costs. This retrospective cohort study aimed to use machine learning methods to develop clinically meaningful models to support clinical decision making.

The primary outcome was the incidence of POD within 3 days postoperatively. The patients will be randomly split into two datasets with split ratios of 80% and 20%.

Subsequently, 80% of the patients will be used for training, and 20% of the patients will be used for testing. Multiple machine learning algorithms will be used to develop POD risk prediction models. The discrimination ability of the prediction models will be assessed by calculating the area under the receiver operating characteristic curve (AUC). The calibration of the model will be evaluated using the Hosmer-Lemeshow goodness of fit test. Decision curve analysis (DCA) will be used to evaluate the net benefits for each threshold probability. The best model will be selected by comparing the performance between the models. Then the SHapley Additive exPlanations (SHAP) will be used to explain the best one.

Inclusion Criteria:

• aged ≥ 65 years
• elective segmentectomy, lobectomy, or esophagectomy surgeries
• general anesthesia

Exclusion Criteria:

• surgery length less than 1 hour
• preoperative cognitive dysfunction
• admission to the intensive care unit
• second operation within 24 hours
• missing data for any variables