Ontology-enhanced Search for Primary Care Medical Literature
Ontology-enhanced Search for Primary Care Medical Literature
Deborah L. McGuinness
Associate Director and Senior Research Scientist, Knowledge Systems Laboratory, Stanford University, Stanford, CA 94305
Primary care physicians are challenged with a vast array of rapidly changing medical literature in a broad spectrum of content areas. It is difficult to stay current with the literature, yet patients’ treatment depends on a physician’s ability to use tests and treatments that are effective medically, of minimum invasiveness, and fit within insurance restrictions (or other financial considerations). This combination strengthens the need for effective coping mechanisms such as online search that supports physicians in finding literature relevant to a patent’s situation and meeting basic quality standards. While standard information retrieval methods provide an excellent foundation, and content-specialized search access (such as Paperchase’s early interface to Medline) may be effective in some cases, many physicians seek smarter and more versatile online search capabilities. We combined techniques in knowledge representation with standard information retrieval tools to produce an ontology-enhanced online search capability for unstructured and semi-structured medical documents. We also addressed the resulting issues of scale and maintenance by building an environment for use by distributed teams of domain-literate, but computer science-naïve users. We will present our ontology-enriched online search environment along with our web tool for collaborative development and maintenance of ontologies.
Introduction
We have built on a history of our work in knowledge representation (KR) [Borgida, et. al., 1989; McGuinness and Patel-Schneider, 1998; McGuinness and Wright, 1998; Brachman, et. al., 1999; Baader, et. al., 1999] and conceptual retrieval [Brachman and McGuinness, 1988] to marry traditional KR elements of semantics and structure web searches. We were motivated by problems AT&T was facing delivering information in broad areas such as medicine for AT&T’s Personal Online Services HealthSite offering and in community information for its Hometown Network offering. The main challenges were to provide usable site navigation tools, user expectation setting, and most importantly, to provide a friendly and usable search function. Our deployed solution included a taxonomy of relevant medical terms for HealthSite and a taxonomy of community related terms for Hometown Network. The taxonomies were used as navigation aids and also in support of query expansion on search terms.
While this is a relatively simple notion, there were a number of considerations in terms of complexity and size of the taxonomy, reasoning infrastructure, interface issues particularly with respect to usability and filtering, and maintenance issues. The initial deployments of search on the two online offerings provided initial solutions, however we explored the space more thoroughly later in our work on FindUR [McGuinness, 1998] in the context of using knowledge representation techniques to support search in semi-structured domains for electronic yellow pages, online calendars, competitive intelligence sites, staffing sites, technical memorandum search applications, etc. Our findings in our previous work were that ontologies could greatly improve recall and not hurt precision (sometimes improving it) when working in circumscribable content areas. We also found that background content taxonomies were useful for site navigation and in user expectation setting. We will describe this work and its evolution in the context of our search work for primary care literature.
We were approached by AT&T Solutions together with Allegheny Hospital in a NIST-funded effort to improve the search experience for primary care physicians who were attempting to find relevant online literature in support of patient treatment. They had previously observed that standard search tools were not meeting physicians’ needs. The needs that we collectively identified were to provide:
1. A form of conceptual retrieval – not depending solely on physician input search terms.
1. Filtering based on article quality and recency. Initially that might include exposing the study method (e.g., randomized control trial, case study, etc.) or publication journal.
1. A summary of important information in an article.
The team had previously begun to address the third point above by hiring medically trained professionals to input “pearls” from articles. These pearls served as a one or two sentence summary of the important findings. While medical professionals were reviewing the documents to generate the pearls, they also ascertained the study method and annotated the documents. Thus, using the tagging to support filtering based on perceived data quality, we could provide a partial solution to the second point above. They also chose the articles for inclusion based on certain content areas where primary care physicians tend to receive more patients. Initial content areas were osteoporosis and lipid disorders. The documents were semi-structured including fields for authors, journal of publication, title, and date.
The team had not begun to address the notion of conceptual retrieval. In this area, they were interested in being able to retrieve articles with related terms. Initial deployment plans were to include only title, author, citation, pearl, and abstract. Thus, it could easily be the case that a short abstract concerning effectiveness studies on dexascans (dual energy x-ray absorptiometry – used as a case confirming diagnostic test for osteoporosis) might not mention osteoporosis in the title, abstract, or pearl. Still a physician who is considering tests for osteoporosis would want to retrieve this article. Similarly, a physician may want to retrieve the same article without needing to know how many ‘x’s are in dexascans and also without having to type in the expanded name of the test.
We attempted to solve the general problem of providing a smarter and customized search solution for the doctors. We also attempted to create an environment in which they could maintain the solution themselves. In the next section, we will introduce our approach to the problem and introduce our tool. Then we will introduce our collaborative tool for maintaining the solution. We will also discuss related work and future directions.
Knowledge-Enhanced Medical Search Interface
We joined the P-CHIP (Primary Care Health Information Provider) team and took responsibility for knowledge representation and search from a design and implementation perspective for the prototype. We intended to rely on our FindUR architecture for search and to extend our tool environment for the maintenance environment.
First we identified our resources:
1. Online abstracts, “pearls”, title, author, date, and journal for documents in the areas of osteoporosis and dyslipidemia.
1. Primary care physician support for knowledge acquisition, structuring, and later, input and maintenance of content information.
1. Support for business case development and analysis.
Next, we identified needs. As noted, users needed a search capability that provided more than typical search engines that take keywords or sometimes natural language (such as Alta Vista or Verity). Also, it is worth noting that our original users were expected to be medical doctors, but later phases were to include medical staff including nurses and physicians assistants, and final stages included patients accessing the system in doctor’s waiting rooms. This broad range of users accentuated the need for support for people who might not be as familiar with medical terms (and their spellings and relations to other medical terms).
When we interviewed physicians, it became clear that the study method of the article was critical. Some doctors wanted to retrieve only documents based on randomized control trial certified evidence. At the very least, they wanted the study method to be quickly obtainable from the document (and preferably, this should be a search criteria). It was also important to be able to search within certain disease areas – for example, it was convenient to search for case confirming procedures only within the osteoporosis literature.
We include our top-level interface for search in Figure 1. It would be obtained after logging into the system. (Logging in allowed us to keep user histories as well as a place for setting user preferences.) The simple version of the search (shown here) allows the physician to function as normal search engines do – allowing input of words or phrases (and then using the embedded Verity search engine to retrieve documents). It also allows a user to select the search within the context of certain known areas in which we had documents. Only the top-level areas are displayed initially, but users may expand the hierarchy underneath any topic area. This screen is taken from an early demo version where we only had content in the areas of osteoporosis and lipid disorders. The current system has eight additional disease content areas. The interface also allows doctors to select by study method.
[pic]
Figure 1- P-CHIP Search Interface
The next figure shows the results of expanding one of the topic areas. The osteoporosis hierarchy has been expanded to show the four main sub-areas under all of the disease categories: diagnostic testing, treatment, epidemiology, and prevention.
[pic]
Figure 2 Expanded P-CHIP Search Interface
After a series of meetings with a number of primary care physicians, it became clear that these were the four areas in which they needed online support. While we did not expand out all of the content areas into these four areas, the doctors believed that it was an organization scheme that could serve generally in their areas. This allowed doctors, for example, to search for all of the articles on hormone replacement therapy for osteoporosis that had a study method of randomized control-trial. Further, they could search for any word in that context, so if they knew they were looking for an article authored by Smith, they could look for Smith in that context. (In this simple interface, we did not let them restrict their searches to look for Smith only within the Author field, but some of our interfaces on other trials included this sophistication as well. Users could also form a query without typing and search, for example for articles on either calcium supplements or calcitonin (or calcium supplements and calcitonin) simply by checking boxes. In repeated user trials, this kind of functionality was desirable. It eliminated typing mistakes and also eliminated spelling errors.
Figure 3 shows the result of a user choosing randomized control trial-certified study method and search for osteoporosis. In this small test, 19 documents matched the query and the first few are presented on the screen. We present the article name, date, and the “pearl” (which was substituted by the first sentence of the abstract when pearls were not available).
[pic]
Figure 3 P-CHIP Search Results I
Figure 3 contained a very simple query and the actual Verity query language was exposed. It however is easy to create much more complicated queries, by choosing Boolean combinations of subtopics off of the hierarchy yet the user never needs to understand the Verity syntax in order to create valid queries. If however, they are experts in the query language, they may insert any arbitrary query in the text box.
The user may click on the title (or the URL) and see the entire listing. Documents display their title, abstract, citation information, and pearl. Also we highlight the reasons why we retrieved the document. For example, a document like “Treatment with Alendronate Prevents Fractures in Women at Highest Risk” may be retrieved as a match to a query for osteoporosis not only because it contains the term osteoporosis, but also because it contains the terms or phrases including alendronate, bone mineral density, BMD, postmenopausal etc. Users can see at a glance all of the relevant terms (and actually can navigate through the document by clicking on an arrow to the next matching term.
Collaborative Ontology Environment
Clearly the background ontologies are the key to the effectiveness of this approach. The domain information was initially collected and organized in a joint team effort between medical doctors and a knowledge representation specialist. After tests proved that the basic organizational structure was sound and after we tested the effectiveness in the original two content areas, then the people who knew the content areas were left in charge of collecting, organizing, and maintaining the content. We customized the collaborative topic builder interface for this domain so that the medical professionals could use it effectively. Figure 5 shows the screen after someone has logged in.
[pic]
Figure 5: Collaborative Topic Builder
The left hand side shows the concepts. Osteoporosis is open and the concept structure below it is shown. The pharmacological treatment area has been opened. On the left, one can see that classes may be added, edited, or deleted (corresponding in the underlying knowledge representation system as modifying subsumption links and concept definitions). On the right side, we can see evidence under a particular node – here pharmacological. One can see evidence phrases immediately under this node like osteoporosis pharmacology as well as evidence for the subclasses of this node such as calcium supplements (and the particular calcium supplements). Users may also modify evidence terms (and they may do this in English or Verity syntax). Figure 6 shows a screen for editing a term.
[pic]
Figure 6: Editing Terms
This frame shows someone editing the drug called Lipitor. The interface supports users in modifying the parent concepts of a term as well as adding or modifying role fillers (that are appropriate for the class of the object). Here we see two known manufacturers for this drug.
Discussion
The P-CHIP project has grown in the last year from two areas to ten, now including diabetes, stroke, asthma, AIDS, pneumonia, hear failure, hypertension, pre-op consults, as well as the original two of osteoporosis and lipid disorders. The basic knowledge organization has held up for the additional eight sub-areas. The underlying structure is still based on Verity with the extensions designed in the FindUR architecture. One could view our system as a knowledge representation-oriented enhancement to Verity’s search and topic environment. FindUR [McGuinness, 1998], the underlying ontology-enhanced search tool marries Verity search, topics, and the CLASSIC knowledge representation system [Borgida, et. al, 1989; Brachman, et. al., 1999]. The actual ontology is loaded from a CLASSIC knowledge base and verified in CLASSIC and then “dumped” into Verity topic form. That allows the more complicated back-end knowledge representation system to run potentially in a standalone environment, while the deployed search system works on mainstream search tools. In this architecture, we may input highly structured and refined ontologies such as those developed in typical description logic-based systems such as CLASSIC. Our current architecture exploits roles (such as manufacturer), value restrictions (such as requiring a filler of the manufacturer role to be a company), cardinalities (such as requiring drugs to have at least one manufacturer and potentially many), and disjointness (such as stating that drugs are not companies).
Using this kind of information, we can identify that inputs stating that one drug is a manufacturer of another drug are erroneous. This kind of error checking is quite valuable when one is trying to maintain high quality and reusable knowledge bases. We found in past applications work (in particular in one family of applications that was in continuous use for a decade) that this kind of early checking for errors eliminated many problems down the line.
We also posit that using a semantically well-grounded underlying representation system provides a much more extensible and understandable framework. It can be useful to know the connection between SERM and Osteoporosis is that the former is a kind of pharmacological treatment for the latter. If later, it was discovered that osteoporosis pharmacological treatments including estrogen were useful as treatments in another disease, then using a knowledge representation system like CLASSIC, the knowledge maintainer may state this in one expression instead of enumerating all of the kinds of treatments for the other disease.
We are not alone in considering description logics for medical applications. The Galen project uses a description logic[Horrocks, et. al., 1996] to represent possibly the largest medical ontology in a description logic-based system. Similarly, K-Rep[Mays, et., al, 1991] has been used to represent medical data as well [Mays, et., al., 1996]. One interesting point for future work would be to see if any of the individually developed knowledge bases would be useful in one of the other applications. Since our approach uses a minimal emphasis on any rules and only exploits the inherent description logic inferences of inheritance, propagation, number restriction, and conflict detection, we may have the best chance of utilizing the other knowledge bases.
Importing other knowledge bases would require two features:
1. Compatible languages or translators
1. Compatible conceptual models or supporting merging tools
Language compatibility within the family of description logics should not be insurmountable. There is a base standard called KRSS – Knowledge Representation System Specification – which resulted from a standards effort supported by DARPA. Our current architecture does not make use of anything that is not captured in KRSS. Also, more recently additional proposed standards have been designed and implemented such as OKBC [Chaudhri, et. al, 1998]. There was an interface to the predecessor of OKBC (called GFP) and there is a wrapper for loading CLASSIC knowledge bases (embodying no more than KRSS) into OKBC-compliant environments such as Ontolingua[Farquhar, 1998].
Convincing others to design knowledge bases in line with one’s preferred conceptual model may be a more difficult issue. We may not be likely to find that others have broken all of their disease categories into subcategories containing testing, treatment, epidemiology, and prevention. We may however find that it is not difficult work to modify their ontologies into this form. Current work such as the Chimaera merging and diagnostic ontology tool at Stanford University[McGuinness, et., al, 1999] begins to address this issue by supporting users in identifying where ontologies might be merged.
Conclusion
We have presented a system for enhancing standard online search with background taxonomies. The resulting system supports physicians in finding relevant documents given a particular patient case. We also have included in the environment, and editing tool that allows domain-literate but possibly computer science-naïve users to maintain the medical content. The initial design was for two content areas, and the current system contains 10 content areas. We propose that the combination of knowledge representation techniques along with standard information retrieval tools for deployment provides a powerful platform on which to build similar systems. Finally, we discussed some other description logic-based systems and suggested issues and solution paths in which the applications may leverage each other.
Acknowledgements We wish to thank our collaborators on the FindUR project – Lori Alperin Resnick, Tom Beattie, Harley Manning, and Steve Solomon. We also wish to thank our many collaborators on the P-CHIP project, notably Russ Maulitz who provided medical expertise and medical project supervision and Ihung Kyle Chang who provided technical support on the medical side. We also thank the medical team who supported us: Eric Vogel, Bob Grealish, and Wes Hutchison and the AT&T Solutions team who supported the project: Nick DiCianni, George Garcia, Chris Sparks, and Sudip Ghatak. We thank AT&T Labs Research for supporting this work and the NIST Advanced Technology Program, which supported the larger project.
Finally, we wish to thank Harry Moore who helped initiate the knowledge-enhanced search work within the AT&T Personal Online Services Division in the context of HealthSite and Hometown Network. We are continuing the work in his memory.
References
1. Baader, F., Borgida, A., Musters, R., and McGuinness, D. L. “Matching in Description Logics”. In Journal of Logic and Computation -- Special Issue on Description Logics. Volume 9, number 3, June 1999.
1. Borgida, A., Brachman, R. J., McGuinness, D. L., and Resnick, L. A. “CLASSIC: A Structural Data Model for Objects''. In Proceedings of the 1989 ACM SIGMOD International Conference on Management of Data, Portland, Oregon, pp. 59--67. June 1989.
1. Ronald J. Brachman, Alex Borgida, Deborah L. McGuinness, and Peter F. Patel-Schneider. "Reducing" CLASSIC to Practice: Knowledge Representation Theory Meets Reality. To appear in the Artificial Intelligence Journal, 1999.
1. Ronald J. Brachman and Deborah L. McGuinness. ``Knowledge Representation, Connectionism, and Conceptual Retrieval'', Proceedings of the 1988 ACM SIGIR International Conference on Research and Development in Information Retrieval, Grenoble, France, June, 1988, pp. 161--174.
1. Chaudhri, V. K., Farquhar, A., Fikes, R., Park, P. D., and Rice, J. P. “OKBC: A Programmatic Foundation for Knowledge Base Interoperability”. In Proceedings of the Fifteenth National Conference on Artificial Intelligence; Madison, Wisconsin; July 26-30, 1998.
1. Farquhar, A., Fikes R., and Rice, J. “The Ontolingua Server: a Tool for Collaborative Ontology Construction”. In Proceedings of Knowledge Acquisition Workshop, Brian Gaines ed. Banff, Canada 1996.
1. Horrocks, I., Rector, A., Goble, C., “A Description Logic Bases Schema for the Classification of Medical Data”, Proceedings of Knowledge Representation Meets Databases - KRDB-96, 1996.
1. Mays, E., Dionne, R., Weida, R., `K-Rep System Overview', SIGART Bulletin, Vol 2(3), pp 93-97, June 1991.
1. Mays, E., Weida, R., Dionne, R., Laker, M., White, B., Liang, C., Oles, F. J., `Scalable and Expressive Medical Terminologies', Proceedings of the 1996 AMIA Fall Symposium, Washington, D.C., 1996.
1. McGuinness, D. L. “Ontological Issues for Knowledge- Enhanced Search''. In the Proceedings of Formal Ontology in Information Systems. June 1998. Also published in Frontiers in Artificial Intelligence and Applications, IOS Press. Washington, DC, 1998.
1. McGuinness, D. L., Fikes, R., Rice, J., and Wilder, S. An Environment for Merging and Testing Large Ontologies. Stanford University Knowledge Systems Laboratory Tech Report KSL-99-05. Submitted for publication, 1999.
1. McGuinness, D. L. and Patel-Schneider, P. F. “Usability Issues in Knowledge Representation Systems''. In Proceedings of the Fifteenth National Conference on Artificial Intelligence, Madison, Wisconsin, July 1998. This is an updated version of “Usability Issues in Description Logic Systems'' published in Proceedings of International Workshop on Description Logics, Gif sur Yvette, (Paris), France, September 1997.
1. Deborah L. McGuinness and Jon Wright. ``Conceptual Modeling for Configuration: A Description Logic-based Approach.'' In the Artificial Intelligence for Engineering Design, Analysis, and Manufacturing Journal - special issue on Configuration, 1998.
................
................
In order to avoid copyright disputes, this page is only a partial summary.
To fulfill the demand for quickly locating and searching documents.
It is intelligent file search solution for home and business.