Locating case discussion segments in recorded medical team meetings

Although there has been great interest in the issue of indexing and providing access to multimedia records of meetings, with substantial efforts directed towards collection and analysis of meeting corpora, most research in this area is based on data collected at research labs, under somewhat artificial conditions. In contrast, this paper focuses on data recorded in a real-world setting where a number of health professionals participate in weekly meetings held as part of the work routines in a major hospital. These meetings have been observed to be highly structured, a fact that is due undoubtedly to the time pressures, as well as communication and dependability constraints characteristic of the context in which the meetings happen. The hypothesis investigated in this paper is that the conversational structure of these meetings enable their segmentation into meaningful sub-units, namely individual patient case discussions, based only on data on the roles of the participants and the duration and sequence of vocalisations. We describe the task of segmenting audio-visual records of multidisciplinary medical team meetings as a topic segmentation task, present a method for automatic segmentation based on a "content-free" representation of conversational structure, and report the results of a series of patient case segmentation experiments. The approach presented here achieves levels of segmentation accuracy (measured in terms of the standard P_k and WD metrics) comparable to those attained by state of the art topic segmentation algorithms based on richer and combined knowledge sources.