Use of Protg-2000 to Encode Clinical Guidelines Ravi D. Shankar, MS, - - PDF document

Use of Protégé-2000 to Encode Clinical Guidelines

Ravi D. Shankar, MS, Samson W. Tu, MS, and Mark A. Musen, MD, PhD

Stanford Medical Informatics, Stanford University School of Medicine, Stanford, California, USA Abstract A major step in building guideline-based clinical care systems is encoding medical knowledge in guideline documents for interpretation by the

• computer. Guideline models provide the structure to

encapsulate guideline knowledge. Developers who are familiar with the guideline model and domain experts who have relevant medical knowledge work as a team to encode guidelines using knowledge acquisition tools. Protégé-2000, developed in our laboratory, is an environment for building knowledge bases. It facilitates knowledge acquisition and maintenance. We have used Protégé-2000 to encode clinical guidelines such as the JNC6 hypertension guidelines and cancer trial

• protocols. Many other modeling groups are also

using it in their work. In this paper, we show that Protégé-2000’s general-purpose knowledge- acquisition framework can provide a rich environment for to encode clinical guidelines. INTRODUCTION A clinical guideline typically contains the guideline’s intentions, medical background, patient eligibility criteria, procedural statements such as clinical algorithms and drug recommendations, evidence for the advisories, treatment cost–benefit analyses, and literature references. Guidelines are generally published in textual format via print and electronic

• media. In recent years, the healthcare community has

shown an increased interest in automated support for guideline-based care. An important step in building such systems is encoding guidelines. This process entails gleaning clinical knowledge out of paper guidelines and using a guideline model to encapsulate the knowledge in a computable

• formalism. The guideline model may be explicit (as in

EON [1]) or implicit (as in PROforma [2]). Knowledge acquisition tools provide a structured environment for domain specialists to build guideline knowledge bases. These tools support the encoding process by providing access to controlled vocabulary and/or user-defined medical concepts, by representing guideline-model concepts, by enabling domain-specialists to enter, edit and browse guideline knowledge via graphical user interfaces (GUIs), and by allowing rapid testing of an evolving knowledge base. Several encoding tools have been developed, and they vary in their architectural flexibility and feature

• support. Some are dedicated authoring environments

for specific guideline models (e.g., Arezzo for PROforma’s guideline model and AsbruView for Asbru [3]). Protégé-2000 [4], on the other hand, is the latest in a series of completely general

• purpose

programs that help users build knowledge acquisition

• systems. Pro tégé-2000’s extensible component-based

architecture and configurable GUI greatly facilitate customizing knowledge-acquisition for given

• domains. Domain experts can use the custom systems

to record, browse and maintain domain knowledge in knowledge bases. P rotégé-2000 is currently used in varied situations from formalizing clinical guidelines to modeling aviation knowledge. Protégé development has been going on for several years, and is still continuing. Today Protégé has evolved to a point where it has become attractive to several guideline-modeling groups (e.g., EON, Prodigy [5], GLIF [6]). This paper illustrates the unique features of Protégé-2000 that strongly support the knowledge acquisition tasks. We discuss the trade-offs using Protégé-2000, a general-purpose framework, to encode guidelines. A BRIEF ANATOMY OF PROTÉGÉ

• 2000

Protégé-2000’s knowledge model [7] is frame-based. A Protégé knowledge base consists of frames that represent classes, slots, facets, instances and

• constraints. Classes are concepts in the domain of
• discourse. They are abstract conceptual entities in

the domain (e.g., the concept of an ACE Inhibitor), or named collections of instances (e.g., sets of Guidelines). Slots are binary -relationships that describe properties or attributes of classes (e.g., the eligibility criteria in the Guideline class). Facets describe properties of slots (e.g., the data type of the eligibility criteria slot). Instances of classes have specific slot values (e.g., in EON, hypertension_guideline is an instance of the Guideline class). Constraints specify additional relationships among instances. A primary feature of Protégé-2000 is its ability to automatically generate a GUI form for each class based on the class definition. Domain experts can use

these forms to enter instances of classes. Besides the forms, complex special-purpose user-interface widgets can also be added to the system as plug-ins to simplify knowledge acquisition (e.g., a diagram widget to encode clinical algorithms). There is also a notion of a Project, which contains configuration information and form layout information. Configuration information includes descriptions of all the user-interface widgets that have been added to the project and a list of other projects that have been included by the current project. ENCODING CLINICAL GUIDELINES A component-based architecture and a facility to expand the core functionality form the foundation of Protégé-2000. Developers can build domain-specific components to solve different knowledge acquisition tasks or simplify them, and to easily integrate them with the system. Thus, rich knowledge acquisition tools custom-tailored for encoding clinical guidelines can be built by assembling, with the core system, special-purpose user-interface widgets, utility functions, and even whole applications. Automatic Generation of User Interfaces When building a knowledge base, most user interaction with the Protégé-2000 GUI is done via forms — users create instances of a class by filling

• ut the form associated with it. Every concrete class

in the knowledge base has an associated form. Protégé-2000 generates user-interface forms automatically based on class definitions, by associating visual metaphors to the data types of the slots in each class, and by using some general layout heuristics (Figure 1). A list of user

• interface widgets

is associated with each data type supported by Protégé-2000. These widgets are visual metaphors that can be used to display the slot values and allow editing them. For example, there is a text field widget to store string slot values, and a check box widget to store Boolean slot values. Sophisticated custom- tailored widgets can also be associated with any data type to generate intuitive user-interfaces. When a form is generated, every slot on a class is automatically associated with a default widget on the

• form. The forms are laid out using simple rules such

as “all slots of similar data types are grouped together”. Developers can use the forms layout editor to override the system’s default presentation by custom tailoring the form’s elements. Users can change the default widget association by choosing a new one from an extensible list of appropriate widgets for each slot type. Users can also modify the default form layout by moving the slot widgets around, by resizing the widgets, and so on. Guideline experts can then use these domain-specific forms to enter instances of the classes, thereby building the guideline knowledge base. Custom User-Interface Widgets Complex domain information such as guideline clinical algorithms warrants specialized visual metaphors for acquiring it and browsing through it. Protégé-2000 allows easy expansion of the standard set of widgets to include custom domain-specific widgets (e.g. a map widget for visualizing geographical information). A complex widget that is useful in encoding guidelines is the diagram widget. The Diagram Widget presents information graphically as a network of nodes and arcs. It comprises a set of nodes with connectors that join

• them. Nodes map to domain objects, and connectors

represent relationships between the objects. Diagram widgets form a natural visual metaphor for encoding a

Figure 2: The Diagram Widget. The hypertension guideline algorithm shown was built using the palette on the right. Figure 1: Automatically generated form for specifying drug information. This GUI was generated based on the Drug_Usage class definition.

guideline’s clinical algorithm (Figure 2). The states, decision points, and action alternatives are modeled as a set of nodes, and the connectors represent a selection among alternatives or a temporal followed- by relation. The diagram itself can be layered, in that a node can represent a sub-diagram. Thus, the clinical algorithm can be visually modeled and presented at different levels of abstraction. The topmost layer depicts a high-level guideline flow, and subsequent layers represent less abstract specifications. Program Plug-ins Protégé-2000’s component-based open architecture facilitates integrating utility functions and custom- built applications into the system. This capability allows developers to tailor the knowledge acquisition environment in a significant way. For example, they can add new functional tabs to the standard set. At knowledge acquisition time, users can access the custom applications via the new tabs. The program plug-ins fall broadly under two categories: the utility functions and end-user applications. Utility functions expand support for knowledge

• acquisition. One functional tab that is relevant to

encoding guidelines allows users online access to the United Medical Language System (UMLS). The UMLS tab allows users to browse online source, to verify the existence of a medical concept within UMLS, and to import sub-trees of the UMLS

• ntology directly into the knowledge acquisition

environment. End-user applications take the ontologies and the knowledge base as input, and can be plugged in as tabs just like the utility functions. The tight integration of the end-user application with our system enhances the functionality of the knowledge acquisition environment immensely. Since changes to the knowledge base are immediately available to the application, they can be tested rapidly by using the application tab. Figure 3. shows a hypertension guideline advisory application as a Protégé-2000 plug-in. This application was built as part of the ATHENA project [8]. Domain experts can rapidly test the advisory system and the entered hypertension guideline knowledge base. They can modify parts of the guideline knowledge base, and immediately see the effects of their changes in the advisories generated by the application. They can also verify the knowledge base against different patient data. Constraint Language Protégé-2000 supports a constraint language called PAL (Protégé Axiom Language), which can be used to write complex integrity constraints on the knowledge base. PAL allows developers to make general assertions about relationships among objects in Protégé-2000 (e.g., “all criteria instances are referenced”, “nodes in a diagram should be

Figure 3: A Hypertension Advisory Application [8] as plug-in tab in Protégé -2000. Using the Knowledge Acquisition tab, domain specialists can enter guideline knowledge. They can test new knowledge base immediately using the application.

connected to other nodes”), and to check if these relationships hold directly in the knowledge base. Another use of PAL is in writing decision-criteria, which define patient-specific constraints that must be evaluated during guideline execution. These criteria are generally comprised of medical-specialty terms such as systolic BP and patient-data terms such as current vitals. An example is “presence of diabetes in problem list and absence of angiotension receptor blocker in authorized medication prescriptions.” This criterion defines Diabetic and not taking ARB. Such relatively simple criteria can be abstracted as classes in Protégé-2000. Guideline authors can then encode specific criteria by filling in corresponding forms generated by Protégé-2000. However, these object templates are not sufficiently expressive to encode complex criteria such as ‘Presence of an authorized medication that is contraindicated by some medical condition.’ This criterion requires t hat, for each current medication, users find its contraindications from the medical- specialty knowledge base, and check for patient-data instances indicating one of these contraindications. Instead of trying to resolve such complex criteria procedurally, developers can use PAL to encode them (Figure 4). PAL implements a subset of first-

• rder predicate logic written in the Knowledge

Interchange Format (KIF) syntax [9]. A PAL structure editor helps developers write these criteria. Projects The component-based architecture extends to the notion of Protégé-2000 projects, too, in that a project can include other projects. This inclusion facility allows building larger custom

• tailored knowledge

bases using reusable smaller knowledge bases. For example, separate knowledge bases can be built to store guideline model concepts and different domain concepts, such as hypertension or cancer domains. Domain-specific custom knowledge acquisition environments can be built by integrating appropriate knowledge bases. Another key feature advanced by the inclusion facility is the notion of integrating multiple guidelines. For example, a guideline on treatment of diabetic patient may include guidelines

• n management of hypertension or foot care.

DISCUSSION Protégé-2000 is a general-purpose knowledge acquisition framework that is widely used by groups in varied fields, inside and outside medical

• informatics. Several guideline modeling groups and

developers of decision-support systems have chosen it as their knowledge acquisition tool. A central question is what the trade-offs are when guidelines are authored using general frameworks such as Protégé-2000, versus using special-purpose tools such as Arezzo and AsbruView. Special-purpose knowledge acquisition tools are tightly coupled with the underlying guideline model. For example, Arezzo supports PROforma, and AsbruView is a user interface for writing plans in the planning language Asbru. Such tools generally provide elegant and sophisticated user-interfaces that are highly directed. Their limitation is that when a guideline model changes, new user interfaces have to be developed, or changes to existing ones have to be made. Making these changes can be complex, costly and time-consuming. Our experience with building guideline-based decision-support systems

• ver the last 15 years has shown that guideline

models evolve during development. Protégé-2000’s core user-interface enables developers to make changes to the underlying model easily. Its automatic user-interface generation facility exposes the new guideline model to the domain-specialists

• immediately. This rapid turnaround is vital for

guideline model evolution and experimentation, and is almost impossible with special-purpose tools. Similarly, combining models may be necessary when building decision-support systems (e.g., combine an

• rganization model with a guideline model). This need

also underscores the advantage of automatically generating knowledge acquisition user-interfaces based on the model changes. On the

• ther

hand, using Protégé-2000’s automatically generated GUI means accepting the basic design choice made by Protégé-2000. For example, Protégé-2000 associates one form with each class and does not facilitate logical grouping of classes into a single form. Therefore, it provides a general forms -based view of guideline knowledge in a knowledge base, but not a concise and domain-

(defrange ?current_med :FRAME Medication) . . . (exists ?current_med (exists ?med_class (and (subclass-of (drug_name ?current_med) ?med_class) (exists ?contraindication (and (Absolute_Contraindications ?med_class ?contraindication)

(exists ?problem

(subclass-of (domain_term ?problem) ?contraindication)))))))

Figure 4. Simplified PAL criteria checking existence of contraindicated medication. As can be seen, the language is expressive but difficult to write in. An editor helps in indenting and matching the parentheses.

specific view. Domain specialists can find it daunting to review the entered knowledge form by form. Application domains often embrace special visual metaphors that provide an intuitive user-interface for entering, editing and browsing domain information. Protégé-2000 promotes expanding the standard set of user-interface widgets and utility functions via a plug-ins facility (e.g. Diagram Widget, PAL editor, UMLS tab). The Protégé group has conducted formal usability experiments that have demonstrated substantial advantages when knowledge acquisition tools provide custom-tailored domain-specific interface components [10]. Changes in the guideline model, medical domain, or user community may require custom user-interface widgets to simplify knowledge acquisition. New utility demands such as access to a medical terminology server will require the expansion of the system’s functionality. Protégé-2000 can tackle these situations by facilitating integration

• f new widgets and utility functions. But the

integration is not going to be as elegant and tight as in special-purpose tools. Therefore, accessing custom functions and using special-purpose widgets can be cumbersome. Protégé-2000 has been used to build decision- support systems based on guideline models with very different goals, such as EON and Prodigy. EON is very expressive and uses complex constructs such as PAL constraints and temporal abstractions to represent complex decision-criteria and patient states. Prodigy is a simpler model that stresses being intuitive to domain-specialists, and relies more on clinicians to recognize complex clinical patterns at the time of guideline execution. Protégé-2000’s plug-and- play framework allowed both the modeling groups to customize the knowledge acquisition environment to suit their models. During encoding a guideline, domain specialists need to be able to share information easily for review, and more often all of them will not have access to, or want to use the knowledge acquisition tool. Protégé-2000 does not support easy review of the guideline

• knowledge. There is no export function to translate

the contents of the knowledge base into a human- readable format. This puts an additional burden on the domain specialists to maintain a textual document

• f the guideline knowledge rules.

Protégé-2000 has a unique infrastructure that makes expanding system functionality easy for developers. Also, its source code is distributed under an open- source license. Thus, it is a fertile platform for different groups to collaborate in enhancing the knowledge acquisition environment. In addition to

• ur team at Stanford, colleagues from around the

world are contributing to the pool of special-purpose plug-ins for Protégé-2000. We expect that with Protégé-2000’s extensible architecture, and with contributions from collaborators including different guideline modeling groups, we can improve the support for encoding guidelines. Acknowledgments This work has been supported, in part, by grant from the U.S. Department of Commerce, National Institute

• f

Standards and Technology, Advanced Technology Program, Cooperative Agreement 70NANB1H3049, by grant from the National Library

• f Medicine LM05708, and by grant from National

Cancer Institute. We thank Valerie Natale for her valuable editorial comments. Protégé-2000 is freely downloadable from our web site under an open source license (see http://protégé.stanford.edu). References [1] Musen MA, Tu SW, Das AK, Shahar Y. EON: A component-based approach to automation of protocol-directed therapy. Journal

• f

the American Medical Informatics Association, 1996: 3(6), 367–388. [2] Fox J, Rahmanzadeh A. Disseminating medical knowledge: the PROforma approach. Artificial Intelligence in Medicine 1998; 14:157-181. [3] Shahar Y, Miksch S, Johnson P. The Asgaard Project: A Task-Specific Framework for the Application and Critiquing of Time-Oriented Clinical Guidelines. Artificial Intelligence in

• Medicine. 1998; 14:29-51.

[4] Musen MA, Fergerson RW, Grosso WE, et al.. Component-Based Support for Building Knowledge-Acquisition Systems. Conference on Intelligent Information Processing (IIP 2000) of the International Federation for Information Processing World Computer Congress (WCC 2000). Beijing, 2000:18-22. [5] Johnson PD, Tu SW, Booth N, Sugden B, Purves IN. Using Scenarios in Chronic Disease Management Guidelines for Primary Care. Proc. of the AMIA Annual Symposium; Los Angeles, 2000; [6] Peleg M, Boxwala AA, Ogunyemi O, et al. GLIF3: The evolution of a guideline representation

• format. Proc. of the AMIA Annual Symposium,

2000; 645-649. [7] Noy NF, Fergerson RW, Musen MA. The Knowledge Model of Protege-2000: Combining Interoperability and Flexibility. 2th International Conference on Knowledge Engineering and Knowledge Management (EKAW 2000). Juan- les-Pins, France, 2000. [8] Goldstein M, Coleman R, Advani A, et al. Implementing practice guidelines for hypertension: effect of computer- generated patient-specific recommendations for drug

• therapy. Medical Decision Making, 1999. 19: 529.

[9] Knowledge Interchange Format: Draft Proposed American National Standard (dpANS). 1998. [10]Noy NF, Grosso W, Musen MA.. Knowledge- Acquisition Interfaces for Domain Experts: An Empirical Evaluation of Protege-2000. 12th Intn’ll

• Conf. on Software Engineering and Knowledge

Engineering (SEKE2000), Chicago, IL2000.