Abstract 11204: Intraoperative Vital Parameters Improve the Accuracy of Machine Learning Based Prediction Model for Perioperative Stroke in Patients With Non Cardiac Surgery

Abstract

Introduction: Perioperative stroke is one of the most devastating events after surgery. To prevent perioperative stroke and stratify the patients at risk, several prediction models or scores based on the preoperative factors were suggested. However, there had been never reported a prediction model using intraoperative physiologic parameters. Aim: We aimed to develop a prediction model for perioperative stroke by analyzing pre- and intraoperative factors using machine learning techniques. Methods: This retrospective cohort study included patients who underwent non-cardiac surgery between 2016 and 2019 at Seoul National University Hospital. Perioperative stroke was defined as a newly developed ischemic lesion at diffusion weighted imaging within 30 days after surgery. Preoperative factors were age, risk factors and laboratory data. Intraoperative variables were blood pressure, heart rate, saturation monitoring value, total amount of fluid and urine volume during surgery. We developed a random forest based prediction model composed of pre- and intraoperative factors and compared with a model consisting of only preoperative features. We validated the model in an external cohort of patients at another hospital between 2020 and 2021. Results: A total of 15752 patients were included in the development cohort, and 109 patients had perioperative stroke. In external validation cohorts, stroke occurred in 11 of 449 patients. The area under the receiver operating characteristics curves (AUC) for integrated models using pre- and intraoperative parameters was significantly higher than that of model using only preoperative factors, as shown in figure(0.822, 95% confidence interval, 0.760-0.883, p <0.001). The AUC of integrated model in external validation cohorts showed the trend in line with that of development cohorts. Conclusions: We demonstrated that using both pre- and intraoperative features may improve the accuracy of prediction for perioperative stroke.