Leveraging Data Science and Informatics in an Automated Detection System of Surgical Errors

点击翻译按钮获取中文摘要

Technological advancements continue to improve surgical outcomes. However, these technologies also introduce new challenges such as communication complexities, equipment troubleshooting under intense pressure, and higher cognitive demand on OR team members. In other words, surgery will continue to be risky despite technological improvements. There is evidence the number of avoidable complications may be underreported, that approximately 39% of in-hospital adverse events are surgical related, and that as many as 4,000 surgical never events (events which should not have occurred) happen in the US each year. The eventual goal of this research is to develop an automated detection system (ADS) of high- risk surgical states. The ADS will prevent surgical safety incidents before they occur through real-time monitoring and notification of appropriate operating room (OR) team members ahead- of-time if there is a looming risk. Thereby allowing the team to reconsider next steps and address the underlying issues, and hence reduce the rates of negative surgical outcomes. This project demonstrates the feasibility and merit of essential components for an ADS. Specifically, the surgical safety literature provides compelling evidence that surgical work-flow disruption (FD) sequences are informative indicators of error causation, therefore it is likely that a future ADS will model and monitor surgical state through tracking flow disruptions. Our current aims are to (1) finish implementation of the Research & Exploratory Analysis Driven Time-data Visualization (READ-TV) research tool; open-source software to visualize FD patterns and other longitudinal data. (2) Develop a stochastic model to predict whether high-risk, disruptive FD sequences will occur based on FD rates at earlier time points. (3) Link FD patterns and sequences with surgical outcomes by developing a text classifier to identify whether or not a surgical safety incident or near-miss occurred based on the associated EHR note. The classifier will be a deep learning model trained with tens of thousands of surgical EHR notes. The text analysis in the third aim will provide insight to FD types and sequences that are more error prone, thereby revealing the FD patterns that an ADS should warn an OR team to avoid. Additional benefits of this text analysis include a possible confirmation of the existence of incident underreporting. Upon completion of the 3 aims, we will have a computational foundation for an ADS: our research tool (aim 1: READ-TV visualization software) and analyses (aim 3: link flow disruptions to safety incidents through EHR note analysis) will advance interpretation of flow disruption (FD) sequences, and our stochastic models (aim 2: predict future surgical state from FD sequences) will prospectively predict error-prone states. This foundation can be extended in future projects through research in automatic transcription of flow disruptions, and the proper mode of alert delivery if the surgery is prone to enter an error-prone state.