HEALTH GeoJunction: place-time-concept browsing of health publications

Abstract

Background: The volume of health science publications is escalating rapidly. Thus, keeping up with developments is becoming harder as is the task of finding important cross-domain connections. When geographic location is a relevant component of research reported in publications, these tasks are more difficult because standard search and indexing facilities have limited or no ability to identify geographic foci in documents. This paper introduces HEALTH GeoJunction, a web application that supports researchers in the task of quickly finding scientific publications that are relevant geographically and temporally as well as thematically.

Results: HEALTH GeoJunction is a geovisual analytics-enabled web application providing: (a) web services using computational reasoning methods to extract place-time-concept information from bibliographic data for documents and (b) visually-enabled place-time-concept query, filtering, and contextualizing tools that apply to both the documents and their extracted content. This paper focuses specifically on strategies for visually-enabled, iterative, facet-like, place-time-concept filtering that allows analysts to quickly drill down to scientific findings of interest in PubMed abstracts and to explore relations among abstracts and extracted concepts in place and time. The approach enables analysts to: find publications without knowing all relevant query parameters, recognize unanticipated geographic relations within and among documents in multiple health domains, identify the thematic emphasis of research targeting particular places, notice changes in concepts over time, and notice changes in places where concepts are emphasized.

Conclusions: PubMed is a database of over 19 million biomedical abstracts and citations maintained by the National Center for Biotechnology Information; achieving quick filtering is an important contribution due to the database size. Including geography in filters is important due to rapidly escalating attention to geographic factors in public health. The implementation of mechanisms for iterative place-time-concept filtering makes it possible to narrow searches efficiently and quickly from thousands of documents to a small subset that meet place-time-concept constraints. Support for a more-like-this query creates the potential to identify unexpected connections across diverse areas of research. Multi-view visualization methods support understanding of the place, time, and concept components of document collections and enable comparison of filtered query results to the full set of publications.

Figure 1

HEALTH GeoJunction System Architecture. The primary components in Health GeoJunction are depicted. As indicated, they are organized as a client-server application supported by a spatially-enabled relational database.

Figure 2

HEALTH GeoJunction initial view. This screen capture shows the default view in GeoJunction (after the user has zoomed in on Southeast Asia on the map) for the sample of documents retrieved from PubMed using a query that identified documents about avian flu and related concepts. The map view represents the frequencies of documents that have been determined to be from or about each country with spilt graduated circles (orange/left side representing documents that are about that country or a place within it and gray/right side representing documents from an organization in that country). Below the map is a double timeline, showing a histogram of document frequency for the entire time period represented by documents (below) and for the currently selected time period (above). In this case, the time period selected is February-November, 2007. The tag cloud at the upper right represents term frequency in the full selected data set. The view shows the two-term pair option. The default order is by frequency. Users have the option to select the more common alphabetical order. The lower tag cloud represents the results of current place, time, and concept filtering. Gray terms have not changed their rank order, green terms have higher rank frequency in the selection than in the overall document set (with bright green representing those terms that are not in the top 100 in the full set), and purple represents terms that dropped in rank. The default color choices support users with color vision deficiencies; custom choices can be set by the user. The space between tag clouds represents currently applied place, time, and concept filters. The bottom right tabbed window provides access to a table of documents, pop-up abstracts, geographic footprints (that display on the map), and a browser that shows the document in PubMed.

Figure 3

GeoJunction view with place and concept filtering. This figure shows the result after the user clicked on the "about" side of the Thailand symbol on the map, filtering the result to only those judged (based on MeSH or GeoJunction feature extraction tools) to be about Thailand, and the user clicked on "disease outbreak" in the top tag cloud to further filter to the subset of documents about Thailand that are also about disease outbreaks. Not surprisingly, H5N1 virus has moved up to be the third most frequent term in this set of documents (now including only 23 of the original set).

Figure 4

Geographic footprints. This figure represents subsequent exploration in which the user has identified two papers of interest. The geographic footprints of both are depicted on the map and the abstract for one is highlighted.

Figure 5

Place, time, concept filter control. This figure shows a detailed view of the facet-based filtering control in which users see the place, time, and concept filters that they have applied. Users are able to selectively remove any of those filters.

Figure 6

Document processing components. The component steps in extracting and geocoding geographic information found in MeSH headings as well as in the title and abstract are delineated here. The approach relies on the OpenCalais named entity extractor to identify geographic references in free text and on the GeoNames database of place names in the world to find geographic entity matches.