Prospective Validation of the Model Predicting Postoperative Delirium Occurrence With Machine Learning-based Analysis of Intraoperative Biological Signals During Anesthesia in Cardiac Surgery

Postoperative delirium (POD) not only increases the length of hospitalization and intensive care unit stay and medical costs, but is also closely associated with negative prognosis, including postoperative mortality, increased morbidity, long-term cognitive decline after surgery, and impaired quality of life and independence. The preoperative risk assessment and early detection of POD are very important in the proper management of POD. This is because drug treatment that can prevent or treat POD is limited, and for its prevention and management, a multidisciplinary approach and resource management covering almost all aspects of patient management are required. Therefore, if there is a model that can predict the occurrence of POD, it can be of great help in managing delirium after cardiac surgery through more accurate risk assessment and early detection. In previous studies, aging and cognitive decline before surgery are known as major risk factors for POD, but identification of risk factors before surgery alone is insufficient to predict the occurrence of POD. Cardiac surgery is highly likely to cause pathophysiological changes that can cause POD, because it is associated with hemodynamic instability, cardiopulmonary use, changes in body temperature, and systemic inflammatory response. These pathophysiological changes can be reflected in the data (biosignals) obtained through various monitoring devices during anesthesia. Most of the events that occur during anesthesia are considered to be correctable risk factors of POD, unlike preoperative risk factors, and there is a potential to reduce the occurrence of POD by actively correcting them. Therefore, it is necessary to analyze the effect of these intraoperative biosignals on POD. In the delirium prediction model development process, rather than simply dividing the already collected data and using it in the model performance validation process, it is better to conduct model performance validation based on patient data prospectively collected to prevent overfitting and achieve higher predictive performance. Therefore, this study aims to collect prospective data to evaluate the performance of the delirium prediction model after cardiac surgery built using machine learning techniques based on the already collected data including biosignals during anesthesia. After reviewing the medical records from the day of surgery to the period of stay in the ICU, if the Intensive Care Delirium Screening Checklist (ICDSC) score is 6 or higher or there is a record of consultation with delirium, it is recorded as POD. After structuring the database through purification, standardization, outlier detection, and sampling of biosignal data generated during surgery, various variables obtained from medical records are collected to construct an evaluation dataset. Using this dataset, the performance of the delirium prediction model built by applying the machine learning algorithm is evaluated through Receiver Operating Characteristic curve analysis.