Machine learning in surgical safety

Post-operative complications have a significant cost, both economically and to the affected individuals. To date, most studies of surgical complications have investigated the pre- and post-operative periods, rather than focusing on the surgery itself. The operating room is a complex environment, and many factors work in tandem. Moreover, assessing surgical quality has been both subjective and time-consuming. Machine learning (ML) can potentially circumvent these identified difficulties. The objective of this program is to use machine learning to identify risk events in surgery that can be used to predict post-operative complications. This program of research focuses on using ML to predict patients at risk for a post-operative complication. This has several subcomponents, including: 1) analysis of the electronic medical record, 2) analysis of surgical video, 3) communication within the team, and and 4) explainable AI. Where possible, efforts will be taken to make inference as close to real-time as possible. This work will use data that is unique worldwide, from the OR Black Box that our group has created, including video recordings of surgery and various statistical covariates, including structured data about the operation, as in the electronic medical record. High risk events on the surgical timeline, including bleeding events, are identified. Standard ML baselines, such as convolutional neural networks, process these data to develop a predictive model for post-operative complications. We already have considerable retrospective data, and the focus is on the algorithms. This proposed study will significantly enhance our understanding of how machine learning can detect surgical factors across speech, text, video, and their combination.