gms | German Medical Science

The Semantic Web is "an extension of the current web in which information is given well-defined meaning, better enabling computers and people to work in cooperation" [Ref. 1]. Whereas the "current web", i.e. the World Wide Web (WWW), provides information primarily for consumption by humans, i.e. natual agents, the Semantic Web allows information to be easily exchanged among natural and artificial agents and to be understood by both [Ref. 2]. The latter is enabled by an extra efforts of humans who make well-defined additions to the information of the WWW. These additions have the form of metadata which can be thought of as annotations to writings or index cards in a library catalogue. In the Semantic Web the structure of these metadata is not arbitrary. It must be such that they can be unambiguously mapped to propositions in a first-order logic (FOL). This requirement is met by committing the authors to a standardized technology for metadata which is called Resource Description Framework (RDF). A proposition is, according to Aristotle, a statement that can be asserted or controverted. Given the successful mapping to FOL, artificial agents are in a position to find out the truth values (true or false) of expressions that are inferred from propositions with known truth values using rules and, therefore, can "understand" the information. The benefit of this mechanism is obvious: In addition to processing information based on syntactical features, artifical agents can perform some reasoning on behalf of their users and complete tasks for them which are more complex than today's syntactical information searches. Examples include validation of proofs as well as retrieval and composition of Web Services. In the following, the first example will be elaborated in more detail. Together with the concept of digital signature the reasoning capability is a necessary (but not sufficient) condition for the transformation of the WWW into a web of trust. It requires that digital signatures be introduced as first class objects, i.e. as globally identifiable objects. Digital signatures can then be part of proofs which are chains of assertions (i.e. asserted propositions) together with inference rules. If an application combines the reasoning engine with the signature verification system it can evaluate whether to trust a given proof or not [Ref. 2]. Not only is it essential for the user to know whether the response of a service is really what was requested and who is the creator of the provided document, but also knowledge about the trustworthiness of an information source is important or - as the cases of ill-advised consumers of medical information show - even critical. This will be discussed below.

Propositional knowledge, i.e. knowledge that can be expressed as propositions, is just one among different kinds of knowledge required to properly apply the SOAP scheme (Subjective, Objective, Assessment, and Plan) at the point of care:The clinical examination starts with the subjective patient history and the current problem as perceived by the patient. These guide the physician in selecting the objective clinical and laboratory examination procedures. On the basis of the subjective and objective information an overall assessment in terms of a diagnosis is made and a treatment plan is established [Ref. 3]. Thus, in addition to the knowledge the physician acquires by asking and examining the patient and consulting the recordings from former visits, she must know how to interpret the test results from the laboratory, know how to apply good clinical practice to the particular case, know where the dosage of drugs and possible adverse effects or drug interactions can be looked up, know how to use the medical terminology for documentation purposes, and, above all, she must have a tacit "sense" for a particular case, which is a function of her personal appraisal of earlier similar cases. Some pieces of the required knowledge can be expressed (and recorded) as propositions, others cannot. Those pieces of knowledge which can be recorded as propositions are usually provided by patient records, test documentation, clinical guidelines, pharmaceutical dictionaries, and medical thesauri. Providing these sources of knowledge in an electronic form is a widely recognized issue challenging medical informatics for a long time. With the Semantic Web this transformation will scale up to a global scope and open the community of "knowledge workers" in charge with the clinical process to artificial agents. Conversely, in the case of experiential knowledge that cannot be made explicit, non-propositional logic, such as fuzzy logic, may be applied to support the clinical process (e.g., for similarity measures in Case-Based Reasoning [Ref. 4]). By default, however, this is the very field where the natural agent - the medical practitioner - brings in all her skills and connoisseurship in order to uniquely frame the particular case in her specific way.

Given the early stage of development of the Semantic Web it is not possible to anticipate all items that should be discussed. However, some points can even be made based on the rough picture painted above. These include (1) the handling of (logical) inconsistency, (2) the notion of trust in the sense of trustworthiness, and (3) the limitations of knowledge representation. Whereas the first issue is addressed by the Semantic Web, dealing with the second is left to the respective application domains, and the third is explored in general Artificial Intelligence. (1) The Semantic Web drops the closed world assumption made by centralized knowledge bases. As anyone can say anything about anything [Ref. 5], the propositions do not have to be globally agreed upon. The cost of the resulting inconsistency is that logical engines cannot be applied to the web as a whole (which would, for reasons of performance, even not be possible in a globally consistent web!). Instead, the engines of the futurewill combine reasoning engines with (syntactical) search engines and (hopefully!) get the best of both [Ref. 6]. Considering, in addition, the process of attributing truth values to propositions introduces the notion of trustworthiness.(2) As mentioned, knowledge about the trustworthiness of an information source (and a given piece of information) is important, particularly in the case of medical information. The provision of this knowledge is addressed by the use of labels and by involving trusted third-parties. Labels are a means for service providers to describe their services. In public health the HIDDEL (Health Information Disclosure, Description and Evaluation Language) metadata vocabulary expressed in XML/RDF provides a model and an element set which can be used for labeling [Ref. 7]. In order to effectively support the clinical process, similar vocabularies are also needed in other relevant medical fields. Trusted third-parties evaluate the labels (and the services) and award seals of approval (i.e. quality marks) to services, such as the MedCERTAIN "transparency mark" in public health [Ref. 7].In order to provide up-to-date knowledge, the quality marks must be reviewed at regular intervals. For the Semantic Web the results of the evaluation must be published in a "machine-understandable" form, preferably by using a common metadata vocabulary. This same vocabulary can also be used by users to describe their preferences. Inconsistency at this level includes possibly conflicting statements made by the service providers and trusted third-parties. In order to solve these conflicts, rules are needed and languages for expressing rules allowing artificial agents to incorporate them for their internal reasoning. The development of reasoning languages is a core issue of the recently established REWERSE European Network of Excellence (Reasoning on the Web with Rules and Semantics) [Ref. 8] which is a follow-up and an extension of the RuleML (Rule Markup Language) initiative [Ref. 9].(3) As mentioned a limitation of the Semantic Web is its inability (or difficulty) to represent something else than propositional knowledge. An implication for the clinical process is that while the artifical agents can support the medical practitioner in performing logical inferences they cannot replace the experience and "intuition" of the natural agent. Thus, the fear that the "autonomous computer" (or computer program such as an artificial agent) is attempting to outsmart the human and that things are getting beyond control [Ref. 10] is not well-founded, at least in the case of the Semantic Web.