Project Director's Narrative of Operational Aspects

Our focus for the first year was on the development of individual components of the system. As described in the summary of research accomplishments, this included research on user interface components, including one for asking questions and infrastructure for integrating results from the various components; search components, including both distributed textual resources and search over video, summarization components, including one for summarization of journal articles, one for summarization of consumer health information, one for automatically building a dictionary of terms and their definitions, and one for summarization of echocardiograms; and components for retrieval of clinical information from the patient records. This work was carried out separately by the different research teams, although we reported regularly on progress in the different teams in biweekly meetings.

To enable integration of the components as well as to foster research on each separately, another main effort was on identifying resources on cardiology for both physicians and patients. In addition to facilitating search over the internet, we are building a local collection of resources for training of the many statistical components of the system, in search, categorization, and summarization. For these purposes, we are building several categories of resources including high quality journal articles, high quality consumer health at various levels of sophistication, and lower quality materials. These will allow us to learn from the categories to automatically partition new material on the internet.

In order to work towards integration of the individual components, we developed a scenario for the system involving a small set of resources and patient records, as described in the summary of research accomplishments. This scenario is serving as the focus for a first test pilot of integrated components. We have collected patient data, journal articles and search requests that are all relevant to the same scenario. We have run experiments with end users, both physicians and residents, to determine how they would ask questions and how they would respond to the same questions that the different system components must respond to. This has yielded a set of gold standards for evaluation of system components.

To prepare for integration, we have begun specifying the format of data that will be passed between the different components. Working with data from the specific scenario has facilitated progress on this end. We have specified interface formats between the search component and the user interface and between the components which extract clinical data and the natural language components that re-rank search results and determine the content of textual summaries.