Synthesizing Intraoperative Multivariate Time Series with Conditional Generative Adversarial Networks

点击翻译按钮获取中文摘要

Project Summary/Abstract Patient safety is paramount in anesthesia. Intraoperative complications and hemodynamic instability are associated with reduced long-term survival and can lead to risks such as myocardial injury, stroke, kidney injury, and even death. Therefore, predicting and preventing intraoperative hemodynamic instability is very important in the decision-making process of anesthesia providers. An ideal pre-operative assessment system would predict, from patient information, all intraoperative complications and physiological changes before a surgical procedure begins. Predicting intraoperative hemodynamic instability during surgery requires analyzing an enormous amount of physiological data and spotting patterns in that data before adverse events occur. However, doing this requires a large volume of high-resolution intraoperative data taken directly from the physiological monitors in the operating room to train machine learning models, and these data currently are unavailable. Therefore, the research goal of this proposed training program is to generate a continuous multivariate intraoperative physiological time series that display the effects of anesthesia management using state-of-the-art mathematic tools. The generated data can provide unlimited and realistic intraoperative data to identify intraoperative complications and later build a real-time intraoperative clinical decision support system. The proposed training program has two aims. Aim 1 will enable the applicant to create a data-driven objective approach for intraoperative complication prediction and risk assessment. Key information from anesthesia pre- op assessment will be used to generate synthetic low-resolution intraoperative physiological data. This data will inform anesthesia providers of the type, timing, and range of a given patient’s intraoperative hemodynamic instability and complications before surgery. Aim 2 will enable the applicant to build a virtual database that will provide unlimited high-resolution intraoperative data to train machine learning algorithms for a future real-time intraoperative clinical decision support system. The recorded low-resolution intraoperative data and the key information from anesthesia pre-op assessment will be inputted into the second tool to upscale existing minute- resolution intraoperative data to second-resolution level for data augmentation to boost the number of available surgical cases. This K08 research program will enable the applicant to fill key knowledge gaps in applying data science in the existing low-resolution intraoperative data in medical records and non-recorded high-resolution intraoperative data displayed by anesthesia devices. The results will orient anesthesia providers and researchers in the design and implementation of data-driven perioperative prediction systems over traditional anesthesia risk assessment. Ultimately, this K08 award will provide the applicant with the senior mentorship, skills, research experience and data resources to become an independent nurse investigator after training.