Early turn-taking prediction in the operating room

This work presents the design and implementation of an early turn-taking prediction algorithm for a robotic scrub nurse system. The turn-taking prediction algorithm analyzes surgeon's implicit communication cues identifying among those surgical instrument requests before the request actually are explicitly evoked. Communication channels expressed through signals like EEG, EMG and physical signs were used to monitor surgeon's behaviors and automatically detect implicit instrument requests. Significant features were extracted from those signals, through an automatic feature selection process. Then recurrent neural networks were used for time-sensitive turn-taking prediction. Experimental results indicated that the proposed algorithm has higher prediction accuracies than human baseline when less than 70% of the entire action was observed. This is approximately 1.4 seconds after the action started, and 0.6 seconds before the action ends. At an extremely early stage (only 10% of data), the proposed turn-taking prediction algorithm achieves a Fl score of 82.8%.