شبکه بیزی سیستم تخصصی پزشکی دینامیکی و مبتنی بر هستی شناسی

The paper proposes an application framework to be used for medicine assisted diagnosis based on ontology and Bayesian Network (DBNO). There are two goals: (1) to separate the domain knowledge from the probabilistic information and (2) to create an intuitive user interface. The framework architecture has three layers: knowledge, uncertainty model and user interface. The contributions of the domain experts are decoupled, the ontology builder will create the domain concepts and relationships focusing on the domain knowledge only. The uncertainty model is Bayesian Network and the probabilities of the variables states are stored in a profile repository. The diagnostician will use the user interface feeded with the domain ontology and one uncertainty profile. The application was tested on a sample medicine model for the diagnose of heart disease.

The diagnosis can be defined as the process of identifying a set of hypotheses that model the problem domain and finding that one with highest probability of matching the real world state. In medical diagnosis, the uncertainty arises from the inability to evaluate the degree of truth of a hypothesis due to unreliable and incomplete information or inconsistent knowledge. The ontology and Bayesian Network (BN) methodologies have been chosen to address knowledge management and uncertainty. The ontology enables the representation of a domain knowledge in a machine understandable form. It can represent the organizational structure of a large complex domains, but the inability to deal with the uncertainty can be a drawback for its application. One disadvantage of the BN is representation of complex structured domains point of view. The ontology and BN can complement themselves in order to overcome the each other disadvantages, thus an ontology-driven uncertainty model can be created. The main goal of the paper is to propose an application framework as a collaborative expert system for creating, developing and maintaining a general model for medicine assisted diagnosis (Fig. 1). From user point of view there will be three roles based on their competencies: concepts and relations definition, connect probabilities to states of the concepts, setting evidences in order to assist the diagnosis. Each role is assigned to one of the triangle’ sides and thus depicting three operational layers. The automated connection between all three layers increases the efficiency of the entire process. The proposed model is implemented as a software using PROTEGE as an ontology framework, NETICA API (Application Programming Interface) as a Bayesian API and Java technology as a development platform. Full-size image (19 K) Fig. 1. Architecture of the proposed concept. Figure options The mapping between domain knowledge and uncertainty model is based on the fact that each concept defined into ontology is part of the BN as a variable. The diagnostician will use an intuitive graphical user interface for changing evidences of the BN variables and based on a threshold the application will depict a chart having the most significant nodes. The final chart will assist the medicine diagnostician to identify the significant factors for a particular case. In order to offer more flexibility the domain knowledge and probabilities are already build and ready to be used as a medicine ontology and respectively uncertainty profile. There is no need to be re-created each time during diagnosis phase. The diagnostician can choose between more than one uncertainty profiles for a medicine ontology. The relation between ontology and uncertainty profile is one to many (in case there are several sources for probabilities tables for the same ontology). This approach speed up the entire diagnosis process and allows the asynchronous update of the ontology and uncertainty profiles by the domain experts in order to increase the degree of accuracy of the information. A similar model, OntoBayes, was proposed in Yi (2007). The major difference between OntoBayes and this proposed model resides in separation between domain knowledge and quantitative component of BN in order to decouple the ontology from the uncertainty probabilities. BayesOWL (Zhongli, 2005) and PR-OWL (Cesar, Costa, Laskey, & Laskey, 2003) are others probabilistic ontology approaches facilitating ontology mapping in the semantic web. Some of their limitations refer to: two-valued variables only allowed in BayesOWL, the ontology model is too complex in PR-OWL. An automated BN construction based on an ontology model was proposed in Devitt, Danev, and Matusikova (2006) to be applied in the telecommunication network management domain. It’s more an algorithm model than an application. Similar approach was proposed in Mihu and Arsene (2009) for monitoring and diagnosis of an IT infrastructure, all layers (gathering evidences, ontology model and presentation layer) were implemented as software agents. Section 2 briefly introduces the ontology concept, the knowledge representation and the implementation of the medicine ontology using PROTEGE application (Horridge, Knublauch, Rector, Stevens, & Wroe, 2004). Section 3 describes Bayesian Network concept, the uncertainty representation and the implementation module within the proposed software application. Section 4 covers the description of proposed software application. Section 5 presents the results of the tests using heart disease data presented in Ghosh and Valtorta, 1999 and Ghosh and Valtorta, 2000.

The DBNO application can be considered as a prototype. The ontology model can be enriched with more medicine diseases, tests, signs in order to have a more accurate diagnosis. A special algorithm was designed for the BN construction and parsing DAG used for graphical representation. From the computation model point of view, the multi-threading Java technology has been used for implementation of the graphical algorithm, thus the parallel computation model gains the advantage of a multi core hardware box. Another advantage is the asynchronous flavor of the three roles activities. Each role is acting independently allowing an information update without blocking the others activity. This application can be used in other domains also; the only constraint is to have at least two classes Cause and Effect. The application is distributed as a Java archive, it can be used very simple on different platforms (Linux, Android, Microsoft) and devices (mobiles, desktops, notebooks, palmtops, servers) which already have a Java virtual machine installed. As further steps the multi-resolution level architecture based on different abstract levels will be considered, thus different levels of information can be accessed instead of having the whole big picture of the domain Dumitrache, Mihu, and Voinescu (2008). The integration with the multi-agent framework JADE (Bellifemine et al., 2007) will be taking into account in order to fulfill the distribution and collaboration purpose of the expert system as well (Dumitrache, Mihu, & Voinescu, 2009).