Hanna Kossowsky Lev

Previous works have shown evidence of upcoming errors prior to their occurrence in EEG data and movement kinematics. To date, the literature has primarily focused on discrete tasks like sequence tapping. In this work, we assess the possibility of predicting upcoming errors in a continuous surgical training task. Surgical residents used the da Vinci surgical robot to move a ring along a curved wire as quickly and accurately as possible. Each collision between the ring and the wire is defined as an error. A group of 18 surgeons completed the task 18 times each, totaling 324 repetitions. We wrote an image processing algorithm to detect the collision errors, yielding labels for our data. Using artificial neural networks, we tested if collision errors could be predicted before they occurred. The input to the networks was segments of kinematic data recorded by the da Vinci. We found that the errors can be predicted with ~75% accuracy, which is well above chance and higher than the prediction accuracy in simpler, discrete tasks. This shows that movement kinematics contain indications of upcoming errors in continuous tasks. This can open the possibility of monitoring kinematics during surgical tasks and alerting for potential upcoming errors.