Development of a guideline authoring tool with PROTÉGÉ II, based on the DILEMMA Generic Protocol and Guideline Model

Size: px

Start display at page:

Download "Development of a guideline authoring tool with PROTÉGÉ II, based on the DILEMMA Generic Protocol and Guideline Model"

Cameron O’Brien’
6 years ago
Views:

1 Development of a guideline authoring tool with PROTÉGÉ II, based on the DILEMMA Generic Protocol and Guideline Model Peter D. Johnson 1 and Mark A. Musen 2 1 PRESTIGE Project c/o Information Department, Royal Brompton Hospital, London SW3 6NP, UK. 2 Section on Medical Informatics, Stanford University School of Medicine Stanford, CA , U.S.A Clinical guidelines are now commonly available, and their benefits have been demonstrated. However, they will not be widely used until they can be freely disseminated in a form that can be applied in the care process to assist the care provider. To do this effectively, guidelines must be authored to a common standard, in a form executable by a computer, such that the guideline takes into account findings already present in the medical record and then offers appropriate guideline-based advice. The Generic Protocol and Guideline Model (GPGM) was developed during the DILEMMA project and aims to solve these problems. However, work in DILEMMA and subsequently on asthma guidelines raises the question as to whether one formal model can adequately represent every piece of knowledge about all guidelines. PROTÉGÉ -II is a suite of programs to aid the development of such knowledge-based systems, which we have used to formally reimplement the GPGM to produce a guideline-editing tool, allowing a clinician to express a guideline as a sharable knowledge base. This process made clear that one model is an over simplification, and a family of related models is required to achieve the goals of the DILEMMA project. INTRODUCTION The benefits of using clinical guidelines and protocols in the practice of medicine have been amply demonstrated[1], but there is a need for a standard computer-based representation. Guidelines can then be shared across many systems using common software to assist the care provider in administering these guidelines. A standard model enables the production of a single common guideline-authoring tool to assist the authoring process. The DILEMMA 1 project focused on decision support in oncology, primary care, and multi-disciplinary care, specifically support for guideline-based care. A product of this project was the Generic Protocol and 1 DILEMMA was a European Community Advanced Informatics in Medicine project Guideline Model (GPGM)[2,3,4] for the representation of the knowledge content of guidelines. PROTÉGÉ-II is a methodology for engineering knowledge-based systems, and a supporting suite of software tools, developed at the Section on Medical Informatics, Stanford University School of Medicine[5]. When used in the domain of protocolbased care[6], PROTÉGÉ-II allows for significant cost savings by generating a protocol-editing tool direct from disparate ontologies[7]. The word Ontology as used in PROTÉGÉ-II is defined as a hierarchical conceptualization of the classes of entities in a given domain, together with their attributes. We have produced a guideline-editing tool to allow clinicians to author guidelines. We used PROTÉGÉ-II to represent the GPGM and create the tool from the resulting ontology. This work has proved a useful case study in model-based knowledge acquisition, demonstrating that a single model to describe all the knowledge content of guidelines is insufficient to represent all uses of such a model, and that a layered approach is needed with models specialized for each use and domain. Generic Protocol and Guideline Model The GPGM was inspired by three earlier projects involving reasoning about clinical decisions the Oxford System of Medicine, Bordeaux Oncology Support System (itself partly inspired by ONCOCIN) and LEMMA[7]. These projects emphasized the benefits of the declarative style of knowledge representation and the separation of domain and control knowledge. The DILEMMA consortium found the initial description of a guideline as a guided sequence of decisions insufficient. More accurately, guidelines and protocols are a representation for producing a care plan for a particular patient, and while this process of producing a plan includes decisions, it is also about actions taken in caring for the patient. Therefore, an existing model of the clinical care process, the Common Basic

2 simplified GPGM (the basic concepts) component protocol has components may be in calculates uses generates from generates State to State transition transition criterion Attribute Result Care plan record (record of what occurred when a guideline applied to the patient) Action State transition made Reason for state transition Attribute value calculated Attribute value used Result value obtained discarded not not required non-existent relevant established scheduled started performed Black arrows show normal life history, gray show exception conditions Arrows show sequence in which transitions can occur. Final states in gray boxes. indicates this transition may occur more than once abandoned has Outcome function Outcome value obtained Figure 2. A subset of valid action state transitions. Figure 1. A Simplified view of the GPGM and how it relates to the patient record. Specification (CBS) model of health care from the United Kingdom National Health Service Information Management Centre [8] was used to supplement the emerging model of guidelines. The CBS model is one of health care as a business process and thus provided a model of agents carrying out activities in organizations and of the resources involved. A clinical variant of the CBS, COSMOS, contributed the notion of action lifecycles[9]. Actions occur over a period of time, and an individual action may progress through a series of recognized states. For example, an infusion may be scheduled at a particular time then started at another time, suspended for a period, restarted, performed, or abandoned. This combination of durations and states is the action s lifecycle. A guideline may be thought of as a plan that consists of many such actions, guided by decisions. The GPGM currently comprises a set of constructs and constraints intended to allow the operational knowledge content of any clinical guideline to be represented in a uniform declarative style. Thus, the GPGM may be considered an ontology for representing the content of protocols and guidelines. The GPGM represents a guideline as a network of interconnected component protocols, which may be composite and hierarchically decomposable. When applied to a particular patient, a component may result in actions being carried out, which are then stored in the patient record. (Figure 1) Not all components result in physical actions; some for example, may be decision actions and others may request data. Actions occur over a period of time, and have a state. The semantics of the model define what comprises a valid state transition. (Figure 2) Each action has criteria for changing from one state to the next. The criteria can be based upon the state of other actions, on data from the patient record, and on information captured during guideline application. Using this overall structure, it is possible to represent the constraints on action sequence and synchronization that are essential for representing guidelines and real clinical processes. Several prototype applications were written by the DILEMMA consortium, using guideline knowledge bases interacting with the care provider and computer-based patient record. Guidelines used included asthma management from the British Thoracic Society, breast-cancer therapy from a tertiary cancer treatment center, angina from a tertiary cardiology center, colorectal cancer, hyperlipidaemia, and several from the Dutch program for primary care protocols. This iteration of authoring, expert validation, and use in the prototypes led to refinements of the GPGM. Four software components are used to apply the GPGM in the DILEMMA architecture. First, the Protocol Task Manager, uses the model-guided knowledge base of the guideline to direct the flow through the guideline, and instantiate the actions to be carried out. Second, the Symbolic Decision Procedure module implements a method for reasoning under uncertainty used to support decision components. Third, a Findings Reasoner abstracts

Figure 3-The GPGM in the PROTÉGÉ-II ontology editor high-level propositions about the current state of the patient from the patient record and entered data.

3 Figure 3-The GPGM in the PROTÉGÉ-II ontology editor high-level propositions about the current state of the patient from the patient record and entered data. The fourth is the application that uses the tools, provides the user interface, and links to the computer-based patient record. There are several intended uses of the GPGM. Such an ontology should define the functionality of the Task Manager component as related to guideline application. The ontology should support the authoring of clinical guidelines by a domain expert through an ontology-based editing tool that enforces the constraints. Such guidelines can be widely disseminated and implemented. The ontology should allow the construction of a single guideline implementation application, which is capable of executing any guideline authored with such a tool, the user-interface requirements being specified by the ontology. Guidelines do not stand alone; development of a standard model is interwoven with other ontologies such as that of the patient record and a standardized terminology. DILEMMA developed an explicit ontology of the computer-based patient record, which was semantically rich enough to enable the GPGM, while still being usable by care providers. There is a very close relationship between the GPGM and the patient-record ontology, each placing demands on the expressibility of the other. For some of the prototypes, terminology was provided by the Read Coding system. PROTÉGÉ-II Developing a knowledge based system in PROTÉGÉ-II involves five steps: (1) Define a domain ontology. This is achieved with the help of the PROTÉGÉ-II ontology editor. Concepts of the domain are entered in a class hierarchy with their attributes. Attributes may be constrained to a fundamental data type, or an instance of another class in the ontology. Other relationships are not part of the ontology, and other constraints are not represented. (2) Construct one or more problem-solving methods, such as the symbolic decision procedure used in DILEMMA, to support decision making under uncertainty in the guideline. (3) Modify the domain ontology to support the problem solving method. For proposing and arguing about solutions required by the symbolic decision procedure, extra slots are required. This is then called the application ontology. (4) Define mapping relations from domain ontology elements to the data inputs required of the problemsolving method. (5) Automatically generate a knowledge-acquisition tool from the application ontology. With the GPGM, this is the guideline authoring tool.

4 METHODS Using the PROTÉGÉ-II ontology editor, we represented the GPGM as an ontology (Figure 3) and built a guideline-authoring tool from the ontology. However, the PROTÉGÉ-II methodology requires more formalization than was available from GPGM documentation. In DILEMMA the GPGM was modeled and used with the language Prolog in a way that allows less formality than the approach of PROTÉGÉ-II. When developing the ontology in PROTÉGÉ-II, all the possible attributes of concepts and relationships between concepts needed to represent a particular guideline must be added at ontology-development time. With the DILEMMA representation, not all possible attributes and relationships need to be modeled in advance; they can be added when required. This means one can use a simplified model and have classes of concepts with varying relationships and attributes. In effect the clinician extends the ontology at knowledge acquisition time, rather than in advance. This flexibility is not possible with the ontology editor approach, so we extended such minimalist classes with subclasses found in DILEMMA guideline knowledge bases. Other difficulties arose because several extended variants of the GPGM have been developed with domain and application-specific extensions, which are required to fully represent a guideline in an executable form. As these were not part of the core defined GPGM they were not entered as part of the ontology. To be able to plan and reason about a guideline, previous actions, diagnoses, investigations, and decisions about the patient must be accessible, together with patient data. These are concepts from the care plan record. During the application of a guideline to a specific case, new entries will be made in the care plan record about decisions taken and actions performed, which may also be used in the planning process. Thus the DILEMMA care plan record ontology was added to the GPGM ontology to enable the representation of all concepts required for guideline knowledge. RESULTS The increased rigor imposed by the PROTÉGÉ-II ontology editor resulted in a representation of the GPGM that required additional subclasses to achieve the same expressibility. This has the benefit of a more formal model, but at the expense of simplicity. However, the complexity is required to achieve a completely model-based representation of a guideline, and hence to be able to author one with a guideline-authoring tool. Domain-specific extensions of the GPGM were developed in DILEMMA. An example comes from the asthma management guideline. A core concept of this guideline is stepped care regimens; where patients are classified into groups based on history and findings and the most appropriate step on a ladder of treatment regimens used[10]. Another area where extension of the model occurred in the DILEMMA project was expressing the knowledge in the guideline needed to control the user interface of the end-user application. For example where the asthma guideline collects data on which to base a decision, the knowledge base defines whether each question has a yes/no answer, a number in a specific range with defined units, or a choice from a list of coded options. The ability to represent such information in the knowledge base is required to meet the goals of a complete knowledge-based representation of the guideline. DISCUSSION Formally representing the GPGM using the more rigorous methodology of PROTÉGÉ-II exposes the single GPGM as a simplification. In reality, the GPGM has evolved into several different layered versions due to the need for additional domainspecific and application-specific concepts. Attempts to bring all extensions back into one common model become unwieldy, especially with the added subclasses required by the more formal representation of PROTÉGÉ-II. A minimal, core ontology is useful for specifying the software component to execute the guideline. Keeping this as simple as possible aids understanding. This core model should be common to all uses of the model to achieve the benefits of a standard guideline representation, particularly those of dissemination. The ontology required to specify the functionality of the Task Manager component is a common subset of the ontology required to define a knowledge-acquisition tool for guidelines, which may require domain- and application-specific extensions. The principle of minimal ontological commitment would suggest that these extensions should be formally developed as separate ontologies. Capturing protocols during DILEMMA showed that the core GPGM may be insufficient as an ontology of more complex protocols, which contain new concepts that much be expressed. An example from the chronic asthma guideline is the concept of stepped care regimens ; a predefined set of treatment regimens with a defined order of use. Naturally, this extension would seem to be part of the process of authoring the guideline, as the realization that an ontology requires extension is often only apparent at that stage. Ideally one would like to be able to extend, but not modify,

5 the basic ontology during the guideline-authoring process. The PRESTIGE: Guidelines in Health Care project is a European AIM project of which one of us is a member, which continues the work from DILEMMA. In this project a guideline-authoring tool is proposed based on such a layered GPGM family. The T-Helper project [6, 7] at Stanford provides another protocol and guideline model that was developed using PROTÉGÉ-II. There are striking similarities between this and the GPGM; both use hierarchically decomposable components, with actions that have duration and state. The main difference is that T-Helper provides a separate algorithm to control the flow through the various actions, whereas the GPGM relies solely on state transition criteria to direct flow. However, one representation should be transferable into the other. In T-Helper only two ontologies were produced; a hybrid domain-specific protocol and patient-record ontology, and a terminology ontology. Subsequent work, however, has also demonstrated the benefit of modularizing these ontologies. CONCLUSION The more rigorous ontology representation of PROTÉGÉ made apparent that one model, the GPGM, is not sufficient for all intended uses. A layered family of ontologies is required based on a common core model. At the core is a minimal guideline model with additional layers for user-interface definition and domain-specific extensions. Domain-extended versions can then evolve independently, within the constraints of their inheritance from the core ontologies. Particular ontologies necessary for guideline authoring, which should be shared from other sources, are those of patient findings, diseases, resources, procedure, and administration terminology. The implementation of computer-based guidelines places heavy demands on the expressibility of the computer-based patient record, so the development of the record needs to be considered along with the development of the core guideline ontology (as in DILEMMA). Ontology development in medicine, then, is not an isolated affair, and one needs a library of ontologies on which to draw. Only when all of these ontologies are available is it possible to create a single, standardized guideline-authoring tool. Acknowledgments Peter Johnson s work on this paper was partially funded by European Commission project HC 1040 PRESTIGE: Guidelines in Healthcare (DGXIII C.4, Telematics Applications for Health), and carried out while a Visiting Scholar at the Section on Medical Informatics, Stanford. Mark Musen is supported by NLM grants LM05157 and LM05304, and by NSF Young Investigator Award IRI We gratefully acknowledge the contributions from DILEMMA and PRESTIGE, especially from Colin Gordon and Ian Herbert. In PROTÉGÉ-II the contributions of Samson Tu and John Gennari were greatly appreciated. References 1. Grimshaw JM, Russel IT. Effect of clinical guidelines on medical practice: A systematic review of rigorous evaluations. Lancet 1993;324: Herbert SIH, Gordon CJ, Jackson-Smale A, Renaud-Salis J-L. Protocols for clinical care. Comp Meth Prog Bio 1995; 48: Gordon C, Herbert SIH, Jackson-Smale A, Renaud-Salis J-L. Care Protocols and Healthcare Informatics. Technology and Informatics 10- Artificial Intelligence in Medicine. IOS Press 1993: Herbert SIH, Gordon C, Jackson-Smale A, Renaud-Salis J-L. Protocols for Clinical Care. Proc 12th Int Congr Euro Fed Med Inf 1994: Musen MA, Gennari J, Eriksson H, Tu SW, Puerta AR. PROTÉGÉ -II: Computer support for development of intelligent systems from libraries of components. Proc World Cong Med Inform (MEDINFO) 1995; Musen MA et al. T-HELPER: Automated support for community-based clinical research. Proc Symp Comp Appl Med Care 1992; Tu SW, Eriksson, H, Gennari J, Shahar Y, Musen MA. Ontology-based configuration of problemsolving methods and generation of knowledgeacquisition tools: Application of PROTEGE-II to protocol-based decision support. Artif Intel Med 1995;7: The Common Basic Specification Generic Model, 3 Vols, NHS IMC, Birmingham, UK The clinical view of the Common Basic Specification: the COSMOS project clinical process model version 2.0. NHS IMC, UK 1993;2(5): Gordon C. Asthma Knowledge Base Design. DILEMMA project report, 1994.

Demonstration of DeGeL: A Clinical-Guidelines Library and Automated Guideline-Support Tools

Demonstration of DeGeL: A Clinical-Guidelines Library and Automated Guideline-Support Tools Avner Hatsek, Ohad Young, Erez Shalom, Yuval Shahar Medical Informatics Research Center Department of Information