Non-local Dependencies and Contextual Information in the Interpretation of Procedural Surgical Text

Abstract

Efforts to automate mining of procedural information from surgical texts are enabling technology for future autonomous surgical robots. In building their knowledge base (or seeking to augment it if they encounter un- expected circumstances), these robots need to interface with surgical resources so far intended for human use— such as textbooks, case reports, published medical lit- erature. Knowledge mined from surgical texts forms the backbone of higher cognition mechanisms that support surgical situation awareness and autonomous decision- making, as well as verbal interaction with humans. Great progress in natural language understanding is required. Regardless of whether the information is in- tended to support monitoring processes, surgeon or performance evaluation, or (in the future) autonomous decision-making, the safety-critical nature of the domain requires a careful study of the information landscape. Surgical textbooks contain descriptions of surgical pro- cedures that are presented semi-algorithmically, often organized in phases and occasionally also in steps. Sur- geons can understand and execute the procedure from the description, so it is possible at least in principle to derive a high-level executable representation of the procedure, one which could be suitable for an autonomous robot. However, even assuming that the robot already have executable routines corresponding to elementary surgical instructions such as incising an anatomical structure, placing an object in an anatomical location (and so on), not all information necessary to parametrize a surgical action is found directly in the text, and when it is, it is not always local to where the instruction is mentioned. Some parameters that cannot be found in the text are not missing per se; rather, they are patient-specific and thus available only when the abstract procedure is instantiated. Their values can only be determined from the direct perception of the surgical scene or from patient-specific imaging.