Real-Time Identification of Operating Room State from Video

Managers of operating rooms (ORs) and of units upstream and downstream of the OR (e.g., postanesthesia care) seek real-time information about OR occupancy to make decisions about managing OR workflow and coordinating resources. Nursing and anesthesia staff typically record patient in/out times by hand, and OR managers spend time walking about the OR suite to estimate the time each case will finish. This thesis describes a system for using real-time video to automatically identify the state of an ongoing operation. This state information is relevant to determining OR occupancy and estimating time to case completion. The system, which uses support vector machines to learn to identify image features relevant to state identification and hidden Markov models to capture the sequential nature of the domain, was tested on video captured over a two day period in one of the nineteen ORs in Baltimore's R. Adams Cowley Shock Trauma Center. An overall accuracy of 99.5% was obtained for identifying for each video frame the corresponding operation state, which was one of the following: OR ready, patient entering, operation in progress, operation ending, patient exiting. Our results are contrasted with the current state-of-the-art system used by the Shock Trauma Center which is based on patient vital signs data.