F:\kathi\conferences\cgp04_\cgp04_timewrap.dvi

Transformation of Structured Guideline Components Institute of Software Technology & Interactive Systems {kaiser, silvia}@asgaard.tuwien.ac.at Abstract. Guideline and protocol representation languages have reached a
level of complexity where auxiliary methods are needed to support the author-
ing of protocols in the particular language. Several approaches and methods
exist that claim high knowledge about both, the medical context and the for-
mal requirements. Therefore, we need knowledge-based methods to facilitate
the human plan designer and create the protocols of the particular language
as automated as possible. We present a three-step wrapper method, called
TimeWrap, to extract information, in particular temporal issues, out of semi-
structured data and integrate it in a formal representation. We illustrate our ap-
proach using the guideline-representation language Asbru and examples from
guidelines to treat conjunctivitis.
Introduction
For better supporting the medical staff during their diagnostic and therapeutic steps, clinicalguidelines and protocols (CGPs) shall proceed in a computer-supported way. Hence, a trans-formation of the CGPs in a (semi-)formal representation that will be executed in an appli-cation is required. Various guideline-representation languages, like Asbru or GLIF (comparethe next section), are available for this reason.
However, clinical guidelines and protocols exist often only in free text. Guideline-repre- sentation languages have accomplished a state of complexity where the generation of suchprotocols is a very challenging venture. As a result we can say that the transformation fromtext to a (semi-)formal representation is mostly either missing or burdensome and time-consuming, but urgently needed to proceed with the task of computer-supported treatmentplanning.
Our aim is to facilitate the generation of computer-supported protocols and in series to support the creation of parts of protocols in Asbru. Asbru is a very complex guideline-representation language and the creation of Asbru protocols is a very sophisticated process.
We have analyzed clinical guidelines to figure out which parts of the guidelines can be usedto easily extract information as automated as possible and convert as well as transform itinto Asbru. Figure 1 illustrates our approach. By means of a domain- and a time ontologyrelevant information is extracted from the clinical guidelines. We are not performing any nat-ural language understanding task to capture the content of guideline components. Afterwards it is integrated into different kinds of intermediate representations and transformed into theformal representation of a guideline-representation language, e.g. Asbru. The application ofintermediate representations is chosen to better structure the content of the CGP and to rep-resent it in a concise form, as e.g. only special aspects, like temporal flows, are represented.
Furthermore a progressive refinement process can be passed through.
ontology repository
ontology
ontology
intermediate
representations
clinical guidelines
(plain text, tables, .)
(semi-)formal representation,
e.g. Asbru
Figure 1: Idea of the method for creating a formal representation of clinical protocols.
We have to pay special attention on temporal aspects of CGPs. To model and to present them in Asbru is a very complex task. It claims both for comprehension of the CGP and goodknowledge about Asbru - especially the representation of temporal flows. On this account wetry to automize the modeling of flows. As a first step we have chosen an area of treatmentplanning: the drug administration. We want to demonstrate this by means of a simple exam-ple.
The next section describes various approaches related to our solution and explains their ben-efits and limitations. In Section 3 we describe requirements regarding the time annotations ofplans and especially cyclical plans in Asbru and in Section 4, we introduce our solution to thesemi-automatic transformation of text to guideline components. We illustrate the usability ofour contribution by a case study in Section 5. Finally we conclude with the discussion of themost important issues and future developments.
Related Work
In the last years various kinds of guideline and protocol representation languages were de-veloped. Thus, the need to support guideline and protocol acquisition and authoring wasemerging and different types of intelligent acquisition methods and tools were developed. Inthe next subsection we illustrate these two development steps.
Guideline and Protocol Representation Languages The major challenges in representing clinical protocols in a computer readable form are toprovide a clear, precise representation with defined semantics and to handle the complexforms of uncertainty which are common in the medical domain. There are several approachesto formalize guidelines or protocols in a computer readable way, e.g., Asgaard/Asbru, GLIF,EON, Prestige, PROforma, Guide (A comprehensive overview can be found at [8]).
Some of these approaches lack a formal definition of their semantics. Often they provide a clearly defined framework but the frames are filled with free text. Such a protocol can there-fore only be interpreted by a human and not by a computer. Also execution or verificationcan only be performed by humans who have to interpret each part of free text and decide itsprecise meaning – an unreliable and often not reproducible process. But there are numerousnotations of logic which provide clear formal semantics. However, the task of modeling a pro-tocol in such a notation is simply impossible to achieve. In particular, intertwined processeswhich develop over time and which involve uncertainty are hard to model in formal logicfrom scratch. The plan-representation language Asbru [7, 10] developed within the Asgaardproject has clearly defined semantics and complex language constructs to represent uncertainand incomplete knowledge.
Guideline and Protocol Acquisition - Intelligent Knowledge Acquisition In the last years, several methods to acquire and extract information from clinical guidelineshave been proposed. Such acquisition tools range from simple editors to sophisticated visualwrappers.
Markup-based tools. Guide-X [12] is a methodology that describes a way to translate
a guideline into a computerized form. An implementation of this methodology was done inStepper [13]. The formalization process is divided into several steps, whereas each step hasan exactly defined input and output.
The GEM Cutter [9] transforms guideline information into the GEM format. It shows the original guideline document together with the corresponding GEM document and makes itpossible to copy text from the guideline to the GEM document. The GEM cutter is similarto our Guideline Markup Tool (GMT) [14], which supports translating guidelines in freetext into the Asbru language, by providing two main features: (i) linking between a textualguideline and its formal representations, and (ii) applying design patterns in the form ofmacros.
These markup-based tools all have in common that the creation process for the comput- erized guidelines has to be done manually by a human plan editor.
Graphic tools. A graphical approach was used in AsbruView [5] which was developed to
facilitate the creation, editing and visualization of guidelines written in the language Asbru.
To be suitable for physicians, AsbruView uses graphical metaphors, such as a running trackand traffic control, to represent Asbru plans.
Two tools are available to translate guidelines into PROforma [4] - both make heavy use of the same graphical symbols representing the four task types in PROforma. AREZZO is de-signed to be used on client-side only, whereas TALLIS [11] supports publishing of PROformaguidelines over the World Wide Web.
These graphic-based tools have in common that they can only be used for design from Wrapper tools. Finally, different kinds of wrappers were developed to transform an
HTML document into an XML document and deliver the extracted data content in XML for- mat with a DTD (for example, XWRAP [6] or LiXto, which provides a visual wrapper [3]).
These methods and tools are very useful in case highly structured HTML documents are used or simple XML files should be extracted. However, clinical protocols are more complexand XML/DTD files that are more structured are needed in order to represent them.
Our approach considers the limitations mentioned above and tries to support the plan genera-tor of guideline components by automating parts of the development process. It is importantto phrase that we are using semi-structured guideline components as source and we are notaiming towards an automatic solution of the transformation process.
In the following section we will explain temporal aspects in Asbru that are required to modelprocesses and that have to be considered in the development of intermediate representationsof processes. In Section 4 we specify our TimeWrap method which tries to overcome thelimitations explained above.
Temporal Aspects in Asbru
Asbru offers extensive possibilities to define complex temporal dependencies and processesby means of Time Annotations. A Time Annotation specifies four points in time relative toa reference point (which can be a specific or abstract point in time or a state transition ofa plan): The earliest starting shift (ESS), latest starting shift (LSS), earliest finishing shift(EFS) and latest finishing shift (LFS). Two durations can also be defined: The minimum du-ration (MinDur) and maximum duration (MaxDur). Together, these data specify the temporalconstraints within which an action must take place (see Figure 2).
Definition:[[ESS, LSS], [EFS, LFS], [MinDur, MaxDur], Reference] Figure 2: Time interval in Asbru. The grey areas indicate the periods when the action has to start andaccordingly finish.
Asbru offers several different types of plans among other things ’cyclical plans’. A cycli- cal plan invokes another plan in consistent periods. For this plan additional temporal annota-tions have to be stated like frequency and possibly the maximum number of cycles. Thereby, the frequency is stated as the period between two iterations that is consistent for the entirecyclical plan (see Figure 3).
cyclical plan
period between the finishing and the beginning The TimeWrap Method
The method we have developed facilitates the extraction of information out of semi-structureddata and integrates the extracted information into a formal representation. This representationis not ultimate. It is a so called ”intermediate representation” capturing the temporal aspects ofa CGP. Other ”intermediate representations” exist that formalize further aspects. Combiningand transforming these parts lead to the definit formal representation – in our case Asbru [10].
Our method takes text – in this example clinical guidelines – as input.
The TimeWrap method consists of three steps: 1. structuring information and representing it in a formal base representation 2. extracting information out of the base representation; and 3. integrating the extracted information into a formal intermediate representation that is the origin for transformation into Asbru. This form of representation can handle tem-poral uncertainties and other demands that are required for planning.
In the following the three steps will be explained in detail.
We have analyzed various clinical guidelines and protocols written in textual form and foundsome typical types of styles.
On the one hand, there always exist diagnosis and treatment parts, which are intertwined and on the other hand, the clinical guidelines are using flow charts and multidimensionaltables to represent diagnostic and therapeutic knowledge. In our first step of analysis, we havechosen therapeutic parts and tables. One very important component of treatment plans is theprescription of drugs. For administering drugs the following information has to be available: • Name of the drug, e.g. Ceftriaxone, Erythromycin, etc.
• Value and unit of the dose, e.g. 1 g, 125 mg or values with composed units like • Kind of application, e.g. orally, intravenous, IV, intramuscular, IM, etc.
• Duration, e.g. 7 days, 10−14 days, etc.
• Frequency of administration, e.g. twice a day, 4 doses a day, etc.
An important part within these definitions for the planning process and in particular for the implementation of Asbru protocols is time-specific data, like the duration of the treatmentand the frequency of the drug-administration.
Most guidelines declare the information about the drug administration by the statement of drug that should be administered and the dosage. The dosage is mostly of the form like’1 g IM, single dose’, ’100 mg orally twice a day for 7 days’, or ’50 mg/kg/day orally dividedinto four doses daily for 10−14 days’.
This information is extracted from tables and integrated in a formal base representation.
The major challange of this step is to cope with a great number of different source formatsand to transform them into a unified format.
Time-specific data and information about the dose rate have to be elicited. This is accom-plished in three steps which are described in the following paragraphs.
(1) Identifying and Processing of Synonyms and Numeric Values For simplifying subsequent processing all expressions that were identified as synonyms areconverted into a consistent expression. As synonyms identified expressions are differentlypresented units, like ’days’, ’day’ or ’milliseconds’, ’msecs’ etc., and numeric expressionswritten in words, like ’single’, ’once’, ’three times’, ’four’. The convertion of numeric ex-pressions into numbers is necessary for subsequent calculations.
(2) Eliciting Data Regarding the Duration and Frequency of Drug Administration The duration should be identified by an expression commencing with numbers followed bya time-unit (e.g. 7 days), or two value or value-unit blocks connected by a dash (–) (e.g. 4 –6 days, 5 days – 2 weeks). The latter describe the duration with a minimum and a maximumlength.
The frequency can be identified by an expression like ’. . . twice a day . . .’, but also by an expression commencing with numbers followed by a time-unit like in ’. . . every 4 hours . . .’.
The latter represents the period between two sequenced actions.
The problem is how to differentiate between two expressions commencing with numbers followed by a time-unit. Which one is the duration? Which one is the frequency? Therefore,we were looking for patterns or methods, which facilitate the differentiation of these expres-sions. We know that the expression specifying the duration must have a greater unit than thefrequency or if the units are both equal the numeric value of the duration has to be greater.
If the expressions were correctly identified as duration and frequency, they are separated into their numeric parts and their unit-parts. If the frequency is stated as ’real’ frequency (e.g.
’twice a day’) it has to be converted into the period between two iterations. That is doneby converting the time-unit into the next smaller time-unit and dividing the new interval bythe number of occurrences. For example ’twice a day’ is first simplified to ’2/(day)’. Thenit is converted to its next smaller unit to ’2/(24 hour)’ and this expression is transformed to’(24 hour)/2’ = ’12 hour’.
One special case appears if a one-time application is prescribed. This is described by the term ’single dose’. In this case we set the value of the duration to ’1’ without stating a unit.
(3) Eliciting Leftover Data Regarding the Dose Rate of the Drug Administration Expressions containing information about the dosage of a drug should contain, like alreadymentioned, the dose rate of the drug, the duration, and the frequency of the administration.
Furthermore, the kind of application and additional information that is not specified any morecan be stated. The sequence of this data may vary and the specification of the duration, thefrequency, the kind of application, and additional information is optional. Hence, the appliedprocedure is the following: We try to mark as many terms as possible besides the already found (duration, frequency).
Then we elicit the dose rate, possibly the kind of application, omitting the duration and fre-quency. The remainding terms, if they are not solely stopwords, are added, too. The resultingterms are combined to the dose rate.
After this step we can generate an intermediate representation that can subsequently be transformed into Asbru. We will describe this task in the following section.
Step 3: Integrating the Extracted Information For the administration of drugs, two types of plans are used that exist in Asbru, too: • A plan that specifies the adminstration of a single dose of the drug. This administration • A plan that is running during a specified period activating a single dose plan in cyclical If neither duration nor frequency is specified in the dosage-expression or ’single dose’ isspecified, only the first plan is used, otherwise both plans are used.
• the frequency of the invocation of the subplan, whereby only the first item is mandatory.
We have defined a schema for this intermediate representation that can represent different types of plans. These plans can be linked together with other plans in sequential or hier-archical order or in an iterative or cyclical order. Additionally, these plans may have timeannotations that may contain uncertainties regarding the begin, the end, and the duration ofthe plan. Time annotations regarding the beginning and the ending are referring to the begin-ning or finishing of another plan that is explicitely stated. It is possible to state multiple timeannotations and different reference plans for the beginning and finishing. In cyclical plansthere is also a declaration regarding the frequency that specifies the time period between thefinishing of the last iteration and the beginning of the subsequent iteration. This is particu-larly important in drug administration, whereby the application in short periods in a row isinhibited.
Case Study
For evaluating our TimeWrap method we used guidelines containing instructions for the ad-ministration of drugs from two different sources. The first guideline is the Preferred PracticePattern (PPP) of the American Academy of Ophthalmology (AAO) for providing guidancefor the pattern of practice for diagnosis and treatment of the patient with conjunctivitis [1].
The second guideline is a Clinical Practice Guideline of the American Optometric Associa-tion (AOA) for the care of patients with conjunctivitis [2].
Both documents contain instructions for drug administration, which are mainly repre- sented in the form of tables. Tables can present data and information in a compressed formmaintaining a concise and structured way. In doing so, a classification of certain data is al-ready comprehensible and concise.
For further processing, the data cannot be used in the available form. It has to be trans- formed into an ”intermediate representation” as shown in Figure 1, in which the informa-tion is also machine-readable. One possibility for such an intermediate representation is thepresentation in XML. At present we are fine-tuning an application that implements an ex-isting method for representing information of a table in a semi-structured way by assigningsemantics. We have obtained an example file for evaluation and testing which is shown inListing 1.
Listing 1: Structured Information: example file about drug administration.
<?xml v e r s i o n = ” 1 . 0 ” encoding =”UTF−8”?>
<!DOCTYPE treatment SYSTEM ” treatment . d t d”>
<treatment>
<cause name=” Gonococcus ” person =” a d u l t ”>
<drug dosage =”1g IM , s i n g l e dose ”
name=” C e f t r i a x o n e ” / >
</ cause>
<cause name=” Chlamydia ” person =” a d u l t ”>
<drug dosage =”100 mg o r a l l y t w i c e a day f o r 7 days ”
name=” D o x y c y c l i n e ” / >
</ cause>
<cause name=” Chlamydia ” person =” c h i l d ”>
<drug dosage =”50 mg / kg / day o r a l l y i n 4 d i v i d e d doses f o r 1 0
name=” E r y t h r o m y c i n b a s e ” / >
</ cause>
<cause name=” Ophthalmia neonatorum ” person =” n e o n a t e ”>
<drug dosage =”25−50 mg / kg IV or IM , s i n g l e dose , n o t t o
name=” C e f t r i a x o n e ” / >
</ cause>
</ treatment>
The discrete entries cover possible classes of dosage indications. The XML-file is parsed and every ’drug’-element is analyzed.
We start with analyzing the value of the dosage-attribute of the first drug-tag. We sim- plify discrete words and detect and convert synonyms into a consistent term. In the presentexpression no synonyms are detected, but the word ’single’ is converted to ’1’. Now we aretrying to elicit the duration, but no numeric value followed by a time-unit is found. The sameapplies for the frequency. The only useful expression found is ’1 dose’ which indicates anonrecurring plan. Therefore, eliciting the dose rate is not necessary, as the complete term fordosage including ’single dose’ is more significant. The resulting intermediate representationis shown in Listing 2.
Listing 2: Intermediate representation for administering a single dose.
<p l a n name=” C e f t r i a x o n e : 1 g IM , s i n g l e dose ”
plan−id =” p l a n 5 5 1 3 1 5 1 2 ” / >
In the second drug-tag, the dosage-attribute contains the value ’100 mg orally twice a day for 7 days’. After identifying synonyms and numeric values written in words the term isconverted to ’100 mg orally 2/ day for 7 day’. The duration is extracted with a value of ’7’ andthe unit ’day’. The frequency is constituted as ’2/ day’ and therefore has to be translated tothe length of the interval between two actions. The time-unit is detected with ’day’, hence thenext smaller time-unit is ’hour’, whereas ’24 hour’ correspond to ’1 day’. The new interval of’24 hour’ is now divided through the number of occurrences ’2’ and thus the result is a valueof ’12’ with the unit ’hour’.
As we have extracted a frequency for the flow of the plan, we can reason on a re-occuring action that is implemented by a cyclical plan shown in Listing 3.
Listing 3: Intermediate representation for administering a drug in cyclical periods.
<p l a n name=” Doxycycline : 1 0 0 mg o r a l l y t w i c e a day f o r 7 days ”
plan−id =” p l a n 5 2 7 6 9 4 4 1 ”>
<c y c l i c a l − p l a n plan−id =” plan5675512”>
<frequency value =”12” u n i t =” hour ” / >
</ c y c l i c a l − p l a n >
<duration>
<min value = ” 7 ” u n i t =” day ” / >
<max value = ” 7 ” u n i t =” day ” / >
</ duration>
<p l a n name=” Doxycycline : 1 0 0 mg o r a l l y ”
plan−id =” p l a n 5 4 6 7 5 5 1 2 ” / >
The third drug-tag contains ’50 mg/kg/day orally in 4 divided doses for 10-14 days’ in the dosage-attribute. We can extract the duration which contains ’10’ as the minimum valueand ’14’ as the maximum value, both with the unit ’day’. We cannot find an expression forthe frequency, as it is covered by the compound unit of the dose rate. Hence, we do not knowthe weight of the person when we generate the plan, we cannot calculate the exact dose rate.
Therefore, we must generate a plan that is specified more precicely during execution.
The compound unit of the dose rate contains the unit ’/day’. Thus we can set the fre- quency to ’1/day’ and can calculate the values and units for the intermediate representation:we convert it into the next smaller unit and get ’1/(24 hour)’ that is then calculated to theperiod between two iterations (’24 hour’). The resulting intermediate representation is shownin Listing 4.
Listing 4: Intermediate representation for administering a drug in cyclical periods.
<p l a n name=” Erythromycin base : 5 0 mg / kg / day o r a l l y i n 4 d i v i d e d
d o s e s f o r 1 0 −14 d a y s ” plan−id =” p l a n 9 7 7 1 2 4 3 1 ”>
<c y c l i c a l − p l a n plan−id =” plan84476443”>
<frequency value =”24” u n i t =” hour ” / >
</ c y c l i c a l − p l a n >
<duration>
<min value =”10” u n i t =” day ” / >
<max value =”14” u n i t =” day ” / >
</ duration>
<p l a n name=” Erythromycin base : 5 0 mg / kg / day o r a l l y i n 4 d i v i d e d
plan−id =” p l a n 8 4 4 7 6 4 4 3 ” / >
The dosage-attribute of the last drug-tag contains ’25-50 mg/kg IV or IM, single dose, not to exceed 125 mg’. Like in the primal tag we find the expression ’single dose’. Thus, wecan reason a one-time application and the resulting intermediate representation is shown inListing 5.
Listing 5: An Intermediate representation for administering a drug in a single dose.
<p l a n name=” C e f t r i a x o n e : 2 5 −50 mg IV or IM , s i n g l e dose , n o t t o
plan−id =” p l a n 5 5 4 9 6 6 3 2 ” / >
After we have finished the generation of the intermediate representation we can transform the data into Asbru plans. Therefore, we created XSLT templates that will do the transforma-tion automatically. Besides templates for cyclical plans we have created templates for plansrelated in a sequential or hierarchical order, too.
By means of an XSLT processor, like e.g. Xalan1, we can generate Asbru plans. The resulting XML-file is valid against the Asbru DTD, but is definitely not a complete Asbruplan. It is a subset representing temporal aspects that can be used within an Asbru protocol(see Listing 6) which has to be further augmented to represent a complete CGP.
Listing 6: Asbru protocol after transforming the intermediate representation.
<?xml v e r s i o n = ” 1 . 0 ” encoding =”UTF−8”?>
<plan−library>
<plans>
<plan−group>
<plan name=” C e f t r i a x o n e : 1 g IM , s i n g l e dose”>
<plan−body>
<user−performed/>
</plan−body>
</plan>
<plan name= ’ Doxycycline : 1 0 0 mg o r a l l y t w i c e a day f o r 7 days
<d e f a u l t s >
<time−annotation>
<time−range>
<duration>
<minimum>
<numerical−constant v a l u e = ” 7 ” u n i t =” day ”/>
</minimum>
<maximum>
<numerical−constant v a l u e = ” 7 ” u n i t =” day ”/>
</maximum>
</ duration>
</time−range>
<now/>
</ time−annotation>
</ d e f a u l t s >
<plan−body>
<c y c l i c a l − p l a n >
<start−time >
1http://xml.apache.org/xalan-j/index.html <time−annotation>
<now/>
</ time−annotation>
</ start−time >
<cyclical−plan−body>
<plan−activation >
<plan−schema name=” Doxycycline : 1 0 0 mg o r a l l y ”/>
</ plan−activation >
</ cyclical−plan−body>
<cyclical−time−annotation >
<time−range/>
<set−of−cyclical−time−points >
<time−point>
<numerical−constant v a l u e =”0”/ >
</time−point>
<o f f s e t >
<numerical−constant v a l u e =”0”/ >
</ o f f s e t >
<frequency>
<numerical−constant v a l u e =”12” u n i t =” hour ”/>
</ frequency>
</ set−of−cyclical−time−points >
</ cyclical−time−annotation >
</ c y c l i c a l − p l a n >
</plan−body>
</plan>
<plan name=” Doxycycline : 1 0 0 mg o r a l l y ”>
<plan−body>
<user−performed/>
</plan−body>
</plan>
<plan name=” Erythromycin base : 5 0 mg / kg / day o r a l l y i n 4
d i v i d e d d o s e s f o r 1 0 −14 d a y s ”> <d e f a u l t s >
<time−annotation>
<time−range>
<duration>
<minimum>
<numerical−constant v a l u e =”10” u n i t =” day ”/>
</minimum>
<maximum>
<numerical−constant v a l u e =”14” u n i t =” day ”/>
</maximum>
</ duration>
</time−range>
<now/>
</ time−annotation>
</ d e f a u l t s >
<plan−body>
<c y c l i c a l − p l a n >
<start−time >
<time−annotation>
<now/>
</ time−annotation>
</ start−time >
<cyclical−plan−body>
<plan−activation >
<plan−schema name=” Erythromycin base : 5 0 mg / kg / day
i n 4 d i v i d e d d o s e s ”/ > </ plan−activation >
</ cyclical−plan−body>
<cyclical−time−annotation >
<set−of−cyclical−time−points >
<time−point>
<numerical−constant v a l u e =”0”/ >
</time−point>
<o f f s e t >
<numerical−constant v a l u e =”0”/ >
</ o f f s e t >
<frequency>
<numerical−constant v a l u e =”24” u n i t =” hour ”/>
</ frequency>
</ set−of−cyclical−time−points >
</ cyclical−time−annotation >
</ c y c l i c a l − p l a n >
</plan−body>
</plan>
<plan name=” Erythromycin base : 5 0 mg / kg / day o r a l l y i n 4
<plan−body>
<user−performed/>
</plan−body>
</plan>
<plan name=” C e f t r i a x o n e : 2 5 −50 mg / kg IV or IM , s i n g l e dose ,
n o t t o e x c e e d 1 2 5 mg”> <plan−body>
<user−performed/>
</plan−body>
</plan>
</plan−group>
</ plans>
</ plan−library>
Results, Benefits, and Limitations
We have shown that by the means of our TimeWrap method time-referenced data of a simpleor cyclical recurring process can be extracted from particular data and out of it planningprocess representations can be created. These processes are first presented in an ’intermediaterepresentation’ and afterwards transformed into a formal language, in our case Asbru.
Thereby, both the often recurring processing of specifications for drug administration and the troublesome generation of Asbru plans can be prevented. Asbru is a very complex lan-guage and not easy to code. Tools that would assist in the process could be very useful. Thus,the knowledge-intensive task of the human plan editor is machine supported, but also theamount of time the process takes can be decreased. By means of the intermediate represen-tation the flows of the clinical protocols can be better structured and presented in a conciseform. The intermediate representation can be used to automatic transform them by definedrules to the final representation Asbru.
Currently, our method handles simple specifications, which are limited by a particular form of information declaration. That means the limitation to one drug that is administeredduring a particular interval in invariant distances of time in a constant dose rate. Besides theselimitations also other dependencies of the administration of drugs like other medications ortreatments or the dependencies of special parameters cannot be processed.
Conclusion
We have presented a three-step wrapper method to analyze and structure semi-structured dataand information that is used to generate a formal representation. We are aiming to support treatment planning within the medical domain and have therefore illustrated our approachwith examples from conjunctivitis and the guideline-representation language Asbru.
It is very important to notice the following issues.
Example of Drug Administration Used As Illustration. The three-step wrapper method presented is illustrated using examples of drug administration. However, TimeWrap can beapplied to similar problem characteristics as well. We have chosen the drug administrationexample because it illustrates our methods more easily.
Automatic vs. Semi-Automatic Transformation. We are not aiming towards an automatic solution to transform different guidelines into formal representations. We are aiming of theautomation of defined semi-structured guidelines’ components, which can interactively becomposed to an overall transformation. However, this last step is done manually.
No Natural Understanding Analysis. We are not performing any natural understanding analysis to capture the content of guidelines. Our starting points are semi-structured guide-lines’ components, which can be processed without syntactic and semantic analysis in thesense of natural language understanding analysis. We definitely need information about thesyntactical shape of the text, but more in a structural sense. Therefore, our methods benefitfrom simplicity, on the one hand, and utilize the known semi-structure forms of the guide-lines’ components, on the other hand.
Our TimeWrap method can be improved according to its ontological foundation. At present, specific expressions and synonyms are defined directly. In the future we will im-plement this by using an ontology.
In the same way the methods for the calculation of the frequency can be improved. The frequent administration of drugs is not distributed equally through the day. In most cases theapplication will be in the daytime. For example the administration three times a day will notbe every eight hours, but perhaps in the morning, at noon and in the evening. On the otherhand, medical domains exist where administration round the clock is necessary.
In the next steps, we will improve our proposed wrapper method and extend the appli- cability to other typical patterns within clinical guidelines and protocols. The overall goalis to design and develop ontology-based wrapper methods, which are applicable to partic-ular classes of knowledge representation, but guided by the idea of clinical guidelines andprotocols.
Acknowledgements
This project is supported by ”Fonds zur F¨orderung der wissenschaftlichen Forschung FWF”(Austrian Science Fund), grant P15467-INF.
References
[1] American Academy of Ophthalmology (AAO): Preferred Practice Patterns Committee.
Cornea/External Disease: Conjunctivitis. American Academy of Ophthalmology (AAO), 1998.
[2] American Optometric Association (AOA): Consensus Panel on Care of the Patient with Conjunc- tivitis. Optometric Clinical Practice Guideline, Care of the Patient with Conjunctivitis. AmericanOptometric Association (AOA), 1995, revised 1999.
[3] R. Baumgartner, S. Flesca, and G. Gottlob. Visual Web Information Extraction with Lixto. In Proceedings of VLDB, 2001. http://www.dbai.tuwien.ac.at/proj/lixto/download.html.
[4] J. Bury, J. Fox, and S. D. The PROforma Guideline Specification Language: Progress and Prospects. In Proceedings of the First European Workshop on Computer-Based Support forClinical Guidelines and Protocols (EGWLP 2000), volume 83: Studies in Health Technologyand Informatics, Leipzig, November 2000.
[5] R. Kosara and S. Miksch. Metaphors of Movement: A Visualization and User Interface for Time-Oriented, Skeletal Plans. Artificial Intelligence in Medicine, Special Issue: InformationVisualization in Medicine, 22(2):111–131, 2001.
[6] L. Liu, C. Pu, and W. Han. XWRAP: An XML-enabled Wrapper Construction System for Web Information Sources. In International Conference on Data Engineering (ICDE), pages 611–621,2000.
[7] S. Miksch, Y. Shahar, W. Horn, C. Popow, F. Paky, and P. Johnson. Time-Oriented Skeletal Plans: Support to Design and Execution. In Fourth European Conference on Planning (ECP’97).
Springer, September 1997.
[8] M. Peleg, S. Tu, J. Bury, P. Ciccarese, J. Fox, R. Greenes, R. Hall, P. Johnson, N. Jones, A. Ku- mar, S. Miksch, S. Quaglini, A. Seyfang, E. Shortliffe, and M. Stefanelli. Comparing Computer-Interpretable Guideline Models: A Case-Study Approach. Journal of the American MedicalInformatics Association (JAMIA), 10(1):52–68, Jan-Feb 2002.
[9] K.-A. Polvani, A. Agrawal, B. Karras, A. Deshpande, and R. Shiffman. GEM Cutter Manual.
Yale Center for Medical Informatics, 2000. http://ycmi.med.yale.edu/GEM/.
[10] A. Seyfang, R. Kosara, and S. Miksch. Asbru 7.3 Reference Manual. Technical Report Asgaard- TR-2002-1, Vienna University of Technology, Institute of Software Technology & InteractiveSystems, Vienna, Austria, Europe, 2002. http://www.asgaard.tuwien.ac.at/asbrusyntax.html.
[11] R. Steele and J. Fox. Tallis PROforma Primer - Introduction to PROforma Language and Soft- ware with Worked Examples. Technical report, Advanced Computation Laboratory, CancerResearch, London, UK, 2002.
[12] V. Sv´atek, T. Kroupa, and M. R˚uˇziˇcka. Guide-X – a step-by-step, markup-based approach to guideline formalisation. In B. Heller, M. L¨offler, M. Musen, and M. Stefanelli, editors, Proceed-ings of the First European Workshop on Computer-Based Support for Clinical Guidelines andProtocols (EWGLP 2000), volume 83: Studies in Health Technology and Informatics, Amster-dam, 2000. IOS Press.
[13] V. Sv´atek and M. R˚uˇziˇcka. Step-By-Step Mark-Up of Medical Guideline Documents. In G. Sur- jan et al., editors, Health Data in the Information Society. Proceedings of Medical InformaticsEurope (MIE2002), pages 591–595, Budapest, HU, 2002. IOS Press.
[14] P. Votruba, S. Miksch, and R. Kosara. Linking clinical guidelines with formal representations. In M. Dojat, E. Keravnou, and P. Barahona, editors, Artificial Intelligence in Medicine. Proceedingsof the 9th Conference on Artificial Intelligence in Medicine in Europe (AIME 2003), Berlin, 2003.
Springer.

Source: http://www.ifs.tuwien.ac.at/~silvia/pub/publications/kai_cgp2004.pdf