The web is becoming a vital tool for the management of medical knowledge. De- ... such as the pocket-sized Oxford Handbook of Clinical Medicine. Despite their ...
Enhancing Conventional Web Content With Intelligent Knowledge Processing Rory Steele, John Fox Cancer Research UK, Advanced Computation Laboratory, Lincoln’s Inn Field, London, England, WC2A 3PX {rory.steele, john.fox}@cancer.org.uk
The Internet has revolutionized the way knowledge can be accessed and presented. However, the explosion of web content that has followed is now producing major difficulties for effective selection and retrieval of information that is relevant for the task in hand. In disseminating clinical guidelines and other knowledge sources in healthcare, for example, it may be desirable to provide a presentation of current knowledge about best practice which is limited to material that is appropriate for the current patient context.. A promising solution to this problem is to augment conventional guideline documents with decision-making and other “intelligent” services tailored to specific needs at the point of care. In this paper we describe how BMJ’s Clinical Evidence, a well-known medical reference on the web, was enhanced with patient data acquisition and decision support services implemented in PROforma.
1 Introduction Health professionals find themselves under increasing pressure from constantly escalating workloads and the growing expectations and demands of patients and managers, with inevitably less time being spent maintaining their personal knowledge bases. Furthermore, the available scientific knowledge that forms the evidence base of everyday clinical practice far exceeds a clinician’s capacity to absorb it and apply it effectively.1 It is widely believed that technologies such as decision support and knowledge management systems have considerable potential to support effective dissemination of up-to-date knowledge to clinicians, bringing relevant information to the right place at the right time, and applying it safely and efficiently (see www.openclinical.org). The web is becoming a vital tool for the management of medical knowledge. Developments in hypertext content have made it possible to rapidly build and publish major repositories of reference information, clinical guidelines and so on. Techniques for automatically generating web content from pre-existing relational- or XMLdatabases, with insertion of links between related sections are also well established. The result is an increasing availability of specialist “knowledge resources” with unprecedented coverage and accessibility. Despite these developments and the impressive performance of modern searchengines, the new web publishing techniques remain problematic. Users who are look1
“Medicine is a humanly impossible task” - Alan Rector
ing for reference documents or seeking answers to questions typically divide their time between navigating across web links and reading the material that seems relevant to their requirements. Although search engines home in quickly on relevant web pages, the end result of a typical search is still a large collection of documents to review in order to find the answers to the original question. A clinician who needs to answer a specific question quickly, or compare treatment options for a particular patient, can still be overwhelmed by material. Healthcare professionals would greatly benefit from search processes that could filter content in a way that focused the presentation on the specific task in hand at the point of care. An alternative technique, developed within the AI community, has been used to control the delivery of content to a user. Some expert-systems have used a set of rules to automatically direct browsing, whilst also providing explanations to the user as to why specific content has been delivered. [1, 2] Unfortunately, translating knowledge from documents into an expert system’s knowledge base has proved difficult. Rada has described a scheme called ‘expertext’ to combine the strengths of both expertsystems and hypertext, [3, 4] with some groups experiencing limited success with this approach. [5] This paper addresses the provision of expert-system like decision support facilities for healthcare professionals, integrated with conventional web content. The aim is to provide patient-specific decision support based on large repositories of web content. The two critical challenges we consider are the need to reduce impractical demands for detailed knowledge engineering, by automatically adapting existing XML content to specific use cases in an expertext-like manner, and the capability to focus presentation of the content in light of the specific clinical needs.
2 Problem description Paper and electronic journals and textbooks are the traditional sources of medical knowledge. Journals normally provide detailed and focused information specific to certain disease-areas, while textbooks typically provide comprehensive information covering aetiology, physiology, diagnosis and treatment. Both formats are normally prepared with quiet study in mind rather than rapid access to patient-specific information. To address the need for rapid clinical reference new print formats have appeared such as the pocket-sized Oxford Handbook of Clinical Medicine. Despite their popularity, such manuals are inevitably lacking in detail and medical publishers continue to look for alternative solutions. One of the most interesting new formats to appear is Clinical Evidence (C.E.), a biannual digest of clinical research developed by the publishers of the British Medical Journal. [6] The basic concept of C.E. is to provide a structured, standardized database of reference information built around (a) major areas of clinical practice, (b) questions that commonly arise about alternative treatments and other interventions in those areas, and (c) the proven benefits - and potential harms - that are associated with different interventions. This structure is illustrated in Figure 1. Each heading in the figure is associated with a certain amount of text, typically in the region of half-a-dozen pages of close print. The user will then need to read the text in order to extract, retain and then correctly apply the evidence provided to make a clinical decision. A web version is also available (www.evidence.org), from which it is possible to drill down from the
evidence summaries into other web resources, notably the PubMed repository of research reports. Despite the popularity of C.E., the staff of BMJ Publishing recognize some important limitations. In its paper form, it is a weighty and unwieldy volume and in its web form, users must navigate up and down a hierarchically tree structure in order to get to the sections they require. More importantly, while the publication provides a uniquely compact review of many areas of modern medical practice, there is still a great deal of information to read and digest. There would still be great value in “filtering and focusing” the content into a form that was directly relevant to specific clinical settings and questions.
Fig. 1. Hierarchical breakdown of Clinical Evidence document structure
Cancer Research UK was asked to carry out an experiment to investigate new ways in which this problem might be addressed. The C.E. knowledge base was supplied to us in the form of a set of XML documents [7] or sections, each containing a text segment about a particular topic (Figure 1). Each topic consists of a set of questions and references. Questions have a set of associated options, with each option describing the benefits, harms and any further comments associated with that option in relation to the posited question.
3 Knowledge authoring The technology used in this experimental system was the Tallis guideline authoring and web-publishing system (www.openclinical.org/kpc). Tallis uses the PROforma process modeling language that was designed to support the specification and execution of task-based processes such as clinical guidelines. PROforma provides an expressive, compositional language based on a small ontology of generic tasks (Fox and Das, 2000): • • •
Decisions - any choice, such as a choice between competing diagnoses or treatments Actions – a simple external act, such as a message action or display of a web page Enquiries – an external request for information, such as a clinical data entry form
•
Plans – any number of the above tasks, possibly including sub-plans
PROforma tasks can be composed into networks, representing processes that are to be carried out over time (such as guidelines, protocols or care pathways). The task specification is a declarative representation that can be interpreted by a suitable engine that enacts the tasks (e.g. acquiring data, evaluating decisions, and controlling the flow of task execution). Task enactment can be influenced by a number of control constructs: • • •
Scheduling-constraints which specify any tasks that must be completed before a task can be considered for execution (e.g. collect data before making a decision) Trigger events which can activate a task independently of any scheduling constraints on it (e.g. user initiates a care pathway) Preconditions that specify any logical circumstances that must hold for a task to be processed by the engine (e.g. a task that should only be enacted if a particular decision has already been made).
A decision task also contains a set of decision options or candidates. Candidates are associated with “argument rules” (which might represents reasons for particular treatments for example) and “commitment” rules, which take or recommend particular candidates based on collections of arguments. Arguments may also be given weightings to indicate that some arguments are “stronger” than others.
4 Integrating a task model into Clinical Evidence The first step in integrating decision support into C.E. was to define a PROforma task structure for the C.E. document structure. This was facilitated by the hierarchical organization of the C.E. document, since all nodes in the C.E. tree structure could map simply to a corresponding PROforma task. As remarked earlier PROforma applications can contain plans that can contain subplans and other tasks. Figure 2 shows how C.E. is modeled as a single plan (represented as a rounded red rectangle), which contains sub-plans that are used as containers for C.E. sections, such as “cardiovascular disease section”. These sub-plans contain the C.E. content dealing with C.E. topics, such as “Acute Myocardial Infarction”. In the C.E. structure topics contain a number of C.E. questions, which are also container plans. The tree structure in Figure 2 also contains many blue squares; each of these represents a PROforma action that contains all the instructions required to display a segment of C.E. text as a web page (including any links to other pages).
Fig. 2. PROforma model based on the content of Clinical Evidence detailed in Figure 1
The next stage in the integration is to “populate” the tasks with knowledge content, the HTML representation of the C.E. content and the knowledge required to provide decision-making. An XSLT [9] document was developed to automatically transform the C.E. XML document into a PROforma document. Within the latter document all the necessary instructions to deliver web pages to the user were encoded as action tasks. For each question within C.E., the XSLT transformation generated a decision support plan, consisting of an enquiry task, to obtain patient data, and a decision task to evaluate arguments for and against different options. The candidates of the decision task were automatically derived from the options associated with the question. The content within the benefits, harms and comment sections on an option was used to create the requisite arguments for a candidate (arguments described within the benefits section “support” a candidate, whilst arguments described within the harms section “oppose” it). Unfortunately, the C.E. text for each option (i.e. the benefits, harms and comments) was not sufficiently well structured to allow arguments to be generated automatically. This process was carried out manually by creating PROforma rules within the Tallis authoring system. [10] The weight of each argument was determined by the strength of the clinical trial data the argument referenced and its statistical significance. To illustrate this manual process, consider the question in Figure 1: “Which treatments improve outcomes in acute myocardial infarction?”. This question has a set of options, one of which is Angiotensin converting enzyme inhibitors (ACE inhibitors). Associated with this option are the following fragments of text: … The overview (4 large RCT s, 98,496 people irrespective of clinical heart failure or left ventricular dysfunction, within 36 h of the onset of symptoms of AMI) compared ACE inhibitors versus placebo. [33] It found that ACE inhibitors significantly reduced mortality after 30 days (7.1% with ACE inhibitors v 7.6% with placebo; RR 0.93, 95% CI 0.89 to 0.98; NNT 200). … The largest benefits of ACE inhibitors in people with AMI are seen when treatment is started within 24 hours.
This would map to the creation of the candidate ‘ACE inhibitors’ for the decision task. This would contain a set of arguments (where ‘ami_onset’ is an integer variable representing the time of infarction onset): •
Argument 1 PROforma argument condition: ami_onset =< 24 PROforma argument weight: +2 PROforma argument caption: Treatment of patients with angiotensin converting enzyme inhibitors within 24 hours of the onset of infarction significantly reduced mortality after 30 days [33]
•
Argument 2 PROforma argument condition: ( ami_onset > 24 ) AND ( ami_onset =< 36 ) PROforma argument weight: +1 PROforma argument caption: Treatment of patients with angiotensin converting enzyme inhibitors within 36 hours of the onset of infarction significantly reduced mortality after 30 days [33]
XSLT tools were also developed to transform the C.E. content into the necessary web pages to provide a façade for the PROforma action, enquiry and decision tasks. Navigational instructions to control guideline enactment, such as task confirmations and triggers, were automatically generated and inserted into the web pages as hyperlinks. Enquiry tasks within the decision support plan can be triggered links within web pages. When activated, a HTML form is provided to query the user for data. The data can then be used by the engine to evaluate which is the most appropriate candidate. Candidates of a decision task can also link to a HTML-encoded breakdown of the relevant arguments, with hyperlinks back to the clinical trials that provided the evidence on which the arguments were based.
5 Application architecture The complete PROforma document and web pages were deployed within a J2EE servlet-container, where they could be enacted via a standard web-browser. In a typical session, the user starts up the PROforma engine, which leads to the activation of an action task that presents all the available sections. The user could then browse to a specific topic, by triggering another action task detailing the questions specific for that topic. For each question, the user may browse through the individual options or activate the decision support facilities, provided as a hyperlink within that web page (Figure 3).
Fig. 3. HTML façade for the Acute Myocardial Infarction Action - showing the decision support hyperlink
Invoking decision support initiates an enquiry task. This displays an HTML form for the user to provide information about the specific patient. On completion of this step the decision task is initiated, using the data collected by the enquiry to evaluate the arguments for and against the different options (candidates). The final task is to construct a report showing the options in order of preference based on an assessment of the overall strength of the arguments. The user can review the arguments for each option.
Fig. 4. HTML façade for the Acute Myocardial Infarction Decision – showing the support for the Angiotension converting enzyme inhibitors candidate
The report in Figure 4 shows a typical report for the decision support service in the application. The top panel shows the 7 decision options for the question “Which treatments improve outcome in acute myocardial infarction?”. The top two ticked options (Nitrates and Blockers) are recommended, while the bottom two crossed options are recommended against. The final choice is made by the user, who may accept the system recommendation or select another option. Here the user has requested further details of one of the equivocal options (indicated by ‘?’) and the arguments are shown
below, one argument “for” and one “against”. The user may also request further justification for the argument, which is provided by linking through to the original research study report located on the PubMed web site. This example application can be accessed for demonstration purposes at http://www.openclinical.org/BMJDemo/demo.html.
5 Discussion and future work The integration of a hypertext document and a PROforma based guideline application has the potential to significantly enhance both the functionality and the usability of traditional web-content. 5.1 Functionality benefits The use of decision tasks allows the user to be directed to the most appropriate content for that current session. Normally, a user would be presented with a web page with a set of hyperlinks to further content. To determine if the information pointed to via these links is relevant, the user is first required to navigate them all, read all the content and then make a decision based on the digested content. With decision support facilities, the doctor merely enters current patient details and the candidate decision options are assessed based upon this data. Such facilities avoid a great deal of unnecessary and time-consuming work by generating only the hyperlinks that are relevant within the particular clinical situation and by ensuring that clinical decisionmaking takes all current information and evidence into account. 5.2 Usability benefits The network structure of a PROforma guideline, and its decomposition into constituent tasks, provides a practical clinical context for information retrieval and navigation. Each content level within the C.E. document maps to a plan within the enactable guideline. Each of these plans then contains further plans (sublevels of C.E. content) and an action task that returns the rendered HTML, containing the necessary links for further navigation. This also provides a context for users with respect to their previous navigational choices. The current set of active tasks can be retrieved at any moment, in effect providing a dynamic set of bookmarks for the current session and reducing time spent browsing back and forth in the complete C.E. document. This could be valuable where time is short, allowing the busy doctor to avoid time-consuming and redundant navigation steps. 5.3 Future work Currently, all HTML content is pre-generated before the guideline is enacted. A future line of enquiry is to generate content dynamically, by providing runtimeprocessing facilities to tasks within PROforma. Such facilities could include the
XSLT generation of the relevant web pages, generation of emails and the querying and/or update of external patient records. Another line of enquiry currently being investigated is to provide a more finegrained description of the content within a C.E. option. Medical ontologies and semantic web initiatives are promising candidates for providing the required high-level descriptions of such content. The automated approach of using XSLT, currently used in the guideline construction, could then be extended to the generation of arguments within the PROforma guideline. The final line of development is to carry out usability testing and clinical evaluation of this approach to decision support. Earlier PROforma applications have been evaluated in collaboration with volunteer doctors in clinical settings. For example, the CAPSULE project found that an “argument-based decision support” system led to a significant improvement in the quality of a doctor’s prescribing decisions (in relation to making a better choice of medication, or a cheaper but equally effective one). [11] The RAGs genetic risk assessment system was evaluated in a clinical simulation and found that decision support technology made sense of the “guideline chaos” in primary care. [12, 13] We aim to carry out a two-stage evaluation of the present technology with paper patients to establish whether (a) decision support of this kind has beneficial effect on clinical decisions, (b) investigate issues of usability and acceptability at the point of care.
6 Conclusion The explosion of medical knowledge on the web is producing problems for the practical retrieval of relevant information at the point of care. A promising solution to this problem is to augment conventional guideline documents with decision-making and other “intelligent” services, tailored to a patient’s specific circumstances. In this paper we have demonstrated how this can be achieved, using BMJ’s Clinical Evidence as the knowledge base and PROforma as the formal guideline representation. The creation of the integrated XML content was partly automatic, with scope for increasing the automated component in such applications, particularly where the target document is well structured. This combination of ordinary documents and formalized knowledge offers a number of potential benefits for improved functionality and usability of guidelines. The present paper has concentrated on the technical aspects of our approach; studies of actual benefits are in progress.
Acknowledgements We would like to thank Dr Jon Fistein and staff of BMJ Publishing for their encouragement in this project and their help in integrating PROforma with Clinical Evidence. We would also like to thank Richard Thomson, Michael Humber, David Sutton and Ali Rahmanzadeh for their help and assistance in use of PROforma and related technologies.
References 1. Shortliffe, E., Scott, A., Bischoff, M., Campbel, A., van Melle, W., Jacobs, C.: ONCOCIN. Expert system for oncology protocol management. Proceedings International Joint Conference Artificial Intelligence. (1981) 876-881 2. Timpka, T.: LIMEDS. Knowledge-based decision support for General Practitioners: an integrated design. Proceedings Tenth Annual Symposium on Computer Applications in Medical Care. (1986) 394-402 3. Rada, R.: Hypertext: from text to expertext, McGraw-Hill, Inc. New York, (1992) 4. Rada, R., Barlow, J.: Expert systems and hypertext. The Knowledge Engineering Review. 3, 4, (1988) 285-301 5. Fox, J.,Glowinski, A., Gordon, C., Hajnal, S., O’Neil, M.: Logic engineering for knowledge engineering: design and implementation of the Oxford System of Medicine. Artificial Intelligence in Medicine. 2, 6, (1990) 323-339 6. BMJ Publishing Group. See http://www.evidence.org 7. World Wide Web Consortium. eXtensible Markup Language (XML) 1.0. W3C Recommendation. See http://www.w3.org/TR/2000/REC-xml-20001006 8. Fox, J., Johns, N., Lyons, C., Rahmanzadeh, A., Thomson, R., Wilson, P.: PROforma - a general technology for clinical decision support systems. Computer Methods and Programs in Biomedicine. 54 (1997) 59-67 9. World Wide Web Consortium. eXtensible Stylesheet Language Trans-formations (XSLT), W3C Recommendation. See http://www.w3.org/TR/xslt 10. Advanced Computation Laboratory. See http://acl.icnet.uk/lab/tallis.html 11. Walton, R.T., Gierl, C., Yudkin, P., Mistry, H., Vessey, M.P., Fox, J.: Evaluation of computer support for prescribing (CAPSULE) using simulated cases. British Medical Journal. 315 (1997) 791-795 12. Coulson, A.S., Glasspool, D.W., Fox, J., Emery J.: RAGs: A novel approach to computerised genetic risk assessment and decision support from pedigrees. Methods of Information in Medicine. 40 (2001) 315-322 13. Emery, J., Walton, R., Coulson, A.S., Glasspool, D.W., Ziebland, S., Fox, J.: Computer support for recording and interpreting family histories of breast and ovarian cancer in primary care (RAGs) - qualitative evaluation with simulated patients. British Medical Journal. 319 (1999) 32-36
