Artificial Intelligence to assist medical experts in collaborative decision making

Lausanne 04 June 1998 Just imagine a number of doctors at different remote locations to dispose of a completely automated discussion forum tool to decide upon the appropriate treatment for a specific patient. No special requirements needed to enter the platform at the exact time as your colleagues do. You can share the experience of a distributed, asynchronous collaboration by means of the assisting guidance of an intelligent system equipped with intuitive interfaces for information retrieval from remote databases and with linguistic modules for natural language processing and argument building. All you need is a Web browser and Internet access to join the medical debate within the user-friendly framework of the World Wide Web.

Advertisement

Just imagine a number of doctors at different remote locations to dispose of a completely automated discussion forum tool to decide upon the appropriate treatment for a specific patient. No special requirements needed to enter the platform at the exact time as your colleagues do. You can share the experience of a distributed, asynchronous collaboration by means of the assisting guidance of an intelligent system equipped with intuitive interfaces for information retrieval from remote databases and with linguistic modules for natural language processing and argument building. All you need is a Web browser and Internet access to join the medical debate within the user-friendly framework of the World Wide Web.

In the Department of Computer Science at the Swiss Federal Institute of Technology, a research team conducted by Nikos Karacapilidis, is working on the development of Hermes. This four component system based on artificial intelligence serves as an assistant and advisor in the complex process of multi-agent medical decision making. The final enforcement of decisions and actions however is left to the agents, who may choose to select a discussion moderator among the participating doctors. The core of the whole architecture is formed by the Argumentation-based Group Decision Support Module, which enables the agents to send various discourse acts to the system's server by electronic mail, in order to present alternative solutions to the discussed medical problem.

Each discourse act is introduced in the system's relational database and the agents interact with it through Java Database Connectivity drivers (JDBC). Hermes distinguishes between four argumentation elements, namely issues, alternatives, positions and constraints which represent preference relations. They are classified in the hierarchical structure of a discussion graph. The issues relate to decisions or goals to be obtained and are introduced by the agents. The alternatives represent the potential choices which correspond to a given issue. The mutual relationship between alternatives can lead to the creation of subissues. In turn, positions consist of statements supporting or resisting the choice for a specific way to handle, expressed in an alternative. As such, they always refer to a single other position or alternative.

Constraints offer the agents a chance to carefully weigh all reasons for and against the selection of a certain modus operandi. They display a typical comparative structure, namely [position, preference relation, position] , in which the preference relation appears to be "more or less important than" or "of equal importance to". The use of constraints provides the alternatives with different levels of importance. Depending both on the argumentation underneath and the specific type of evidence, all argumentation elements but the issues, receive an activation label to indicate their current status, that can be whether active or inactive. A status change is automatically propagated upwards. In case an alternative is affected, the issue has to be adapted according to the new point of view.

The Hermes system uses three levels of proof standard for the elements in the argumentation process. "Scintilla of Evidence" (SoE) applies if at least one active position argues in favour of a given position. If there are no active positions speaking against a position taken in consideration, the latter is "Beyond Reasonable Doubt" (BRD). When the active positions supporting a specific position outweigh those speaking against it, the agents are dealing with "Preponderance of Evidence" (PoE). PoE thus can supply an alternative with a positive activation label if there are no other alternatives with a larger score in the same issue. In a similar way, the activation label of the constraints is determined by the underlying discussion as well as by the activation label of their constitutive positions.

In addition, constraints also obtain a consistency label, which makes them whether consistent or inconsistent. The system compares both positions of a new constraint with those of the previously inserted ones. If they already exist, the new constraint is redundant, in case it has the same preference relation, or conflicting if not. Redundancy is ignored whereas a conflicting constraint is integrated with its counterpart in a new issue to be discussed by the agents until only one of them becomes active. If the new constraint's positions cannot be found in the discussion graph, consistency is compared with previous active as well as consistent constraints, integrated in the same issue. The comparison process is executed by a powerful polynomial path consistency algorithm.

Since Hermes is a web-based system, the agents simultaneously can consult all multimedia information relevant to the argumentation item, highlighted in the upper windowpane of the discussion forum, in the corresponding lower pane by clicking on the URL entry. The second component constitutes the Natural Language Processing Module (NLP). NLP deals with the contents analysis of the agents' messages in providing the appropriate techniques to automatically extract and link the corresponding discussion items. The procedure includes three steps, namely element extraction to segment the message; theme identification for conceptual association; and attitude link-up to display elements and relations in the discussion graph. For this matter, off-the-shelf segmenters, taggers, broad-coverage grammars, and probabilistic partial parsers are being implemented.

The Information Retrieval Module allows the doctors to use various types of information stored in remote databases during discussion. The data can be accessed through the CORBA (Common Object Request Broker Architecture) communication protocol, which enables interoperability in heterogeneous computing environments. The fourth component, referred to as the Argument Builder Module, acts as an advisor offering perusal in the discussion's patterns. As such, the agents are suggested to perform discourse acts which best reflect their interests and intentions. The module is currently under development, in order to provide evaluation of the introduced arguments by means of a set of grammar and syntax rules.

Researcher Karacapilidis believes Hermes to be a promising tool for medical decision making, since it offers a feasible solution to integrate the subtlety of human argumentation with the logic of machine reasoning. Of course, the process of computed collaborative decision making will always remain fallible but still it is able to provide participants with a better understanding of their mutual arguments. The basis of any decision process are the choice criteria. They need to be clearly defined as well as the underlying motives for their selection. If not, the quality of the decision process may suffer from possible implicit frustration between participants of unequal stature, relating to the existence of hierarchical relations. You can find a full description of the Hermes system at the Web site of the Swiss Federal Institute of Technology.


Leslie Versweyveld

[Medical IT News][Calendar][Virtual Medical Worlds Community][News on Advanced IT]